Innovative Approach to Estimating Demand for Intercity Bus Services in a Rural Environment

December 2016

Prepared for US DOT

Prepared by

Jeremy Mattson North Dakota State University Upper Great Plains Transportation Institute Small Urban and Rural Transit Center

Disclaimer

The contents of this report reflect the views of the authors, who are responsible for the facts and the accuracy of the information presented herein. This document is disseminated under the sponsorship of the Department of Transportation University Transportation Centers Program and the Florida Department of Transportation, in the interest of information exchange. The U.S. Government and the Florida Department of Transportation assume no liability for the contents or use thereof.

The opinions, findings, and conclusions expressed in this publication are those of the authors and not necessarily those of the State of Florida Department of Transportation.

Innovative Approach to Estimating Demand for Intercity Bus Services in a Rural Environment

Prepared for:

U.S. Department of Transportation

Prepared by:

Jeremy Mattson

Small Urban and Rural Transit Center Upper Great Plains Transportation Institute North Dakota State University Fargo, ND

Final Report

2177060-NCTR-NDSU10

December 2016

National Center for Transit Research

A USDOT Transit-focused University Transportation Center consortium led by University of South Florida

4202 E. Fowler Avenue, CUT100, Tampa FL 33620-5375 www.nctr.usf.edu

Member universities: University of South Florida, North Dakota State University, University of Illinois at Chicago, Florida International University Metric Conversion

SYMBOL WHEN YOU KNOW MULTIPLY BY TO FIND SYMBOL

LENGTH

in inches 25.4 millimeters mm

ft. feet 0.305 meters m

yd. yards 0.914 meters m

mi miles 1.61 kilometers km

VOLUME

fl. oz. fluid ounces 29.57 milliliters mL

gal gallons 3.785 liters L

ft3 cubic feet 0.028 cubic meters m3

yd3 cubic yards 0.765 cubic meters m3

NOTE: volumes greater than 1000 L shall be shown in m3

MASS

oz. ounces 28.35 grams g

lb. pounds 0.454 kilograms kg

megagrams T Short tons (2000 lb.) 0.907 Mg (or "t") (or "metric ton")

TEMPERATURE (exact degrees)

5 (F-32)/9 oF Fahrenheit Celsius oC or (F-32)/1.8 Technical Report Documentation 1. Report No. 2. Government Accession No. 3. Recipient's Catalog No.

2177060-NCTR-NDSU10 4. Title and Subtitle 5. Report Date

Innovative Approach to Estimating Demand for Intercity Bus Services December 2016 in a Rural Environment 6. Performing Organization Code

7. Author(s) 8. Performing Organization Report No. Jeremy Mattson

9. Performing Organization Name and Address 10. Work Unit No. (TRAIS)

Small Urban and Rural Transit Center Upper Great Plains Transportation Institute 11. Contract or Grant No. North Dakota State University Fargo, ND 2177060-NCTR-NDSU10 12. Sponsoring Agency Name and Address 13. Type of Report and Period Covered

National Center for Transit Research Center for Urban Transportation Research University of South Florida 14. Sponsoring Agency Code 4202 East Fowler Avenue, CUT100, Tampa, FL 33620-5375

15. Supplementary Notes

16. Abstract

Because existing models have their limitations, there is a significant need for a model to estimate demand for intercity bus services, especially in rural areas. The general objective of this research was to develop an intercity mode choice model that can be incorporated into a statewide travel demand model to estimate demand for rural intercity bus services. Four intercity transportation modes were considered in the study: automobile, bus, rail, and air. A stated preference survey was conducted of individuals across the state of North Dakota, and a mixed logit model was developed to estimate a mode choice model. Results from the mode choice model showed the significant impacts of individual, trip, and mode characteristics on choice of mode. Gender, age, income, disability, trip purpose, party size, travel time, travel cost, and access distance were all found to have significant impacts on mode choice, and traveler attitudes were also found to be important. The study demonstrated how the mode choice model can be incorporated into a statewide travel demand model, and intercity bus mode shares were estimated for origin-destination pairs within the state. Alternative scenarios were analyzed to show how mode shares would change under different conditions or service characteristics. This study was conducted in the largely rural state of North Dakota, but results could be transferable to other areas with similar geographic characteristics.

17. Key Words 18. Distribution Statement

Intercity bus services, travel demand model, mode choice 19. Security Classif. (of this report) 20. Security Classif. (of this page) 21. No. of Pages 22. Price Unclassified Unclassified 149 TABLE OF CONTENTS

1. INTRODUCTION ...... 1 1.1 Research Objectives ...... 1 1.2 Organization ...... 2 2. LITERATURE REVIEW...... 3 2.1 Basic Structure of Existing Models ...... 3 2.1.1 Major Ridership Generators ...... 3 2.1.2 Characteristics of Population ...... 3 2.2 Route, Corridor, and City-Pair Models ...... 3 2.2.1 Overview of Early Research Efforts ...... 4 2.2.2 Route-Level Research for Rural Intercity Bus Service ...... 5 2.2.3 Other Studies ...... 6 2.3 Intercity Bus Network Model ...... 7 2.4 Disaggregate vs. Aggregate Models ...... 7 2.5 Mode Choice Models ...... 8 2.6 Statewide Travel Demand Models ...... 9 2.7 Other Approaches ...... 10 2.8 Conclusions ...... 10 3. METHODS ...... 12 3.1 Survey Development ...... 12 3.1.1 Demographic Characteristics ...... 12 3.1.2 Stated Preference Survey ...... 12 3.1.3 Attitudinal Questions ...... 18 3.2 Survey Administration ...... 18 3.2.1 Study Population, Sample Frame, and Drawing the Sample ...... 18 3.2.2 Sample Size ...... 20 3.3.3 Survey Distribution ...... 20 3.3 Discrete Choice Modeling ...... 20 3.3.1 Multinomial Logit Model ...... 21 3.3.2 Nested Logit Model ...... 22 3.3.3 Mixed Logit Model ...... 23 3.3.4 Determining Appropriate Model ...... 24 3.4 Statewide Travel Demand Model ...... 24 4. SURVEY RESPONSE AND SUMMARY STATISTICS ...... 25 4.1 Response Rate ...... 25 4.2 Summary Statistics ...... 25 4.2.1 Demographic Characteristics of Respondents...... 25 4.2.2 Transportation Characteristics of Respondents ...... 28 4.2.3 Opinions on Ease of Travel and Need for Improvements ...... 30 4.2.4 Opinions on Travel ...... 31 4.2.5 Stated Preference Response ...... 34 5. MODEL SPECIFICATION ...... 39 5.1 Mixed Logit Mode Choice Model ...... 39 5.2 Binary Logit Models to Estimate Effects of Attitudes on Mode Choice ...... 42 6. MODE CHOICE MODEL RESULTS ...... 44 6.1 Mixed Logit Model Results ...... 44 6.1.1 Impacts of Individual Characteristics ...... 46 6.1.2 Impacts of Trip Characteristics ...... 46 6.1.3 Impacts of Mode Characteristics ...... 46 6.2 Attitudes and Mode Choice Results ...... 47 7. APPLYING THE RESULTS FROM THE MODE CHOICE MODEL ...... 50 7.1 North Dakota Statewide Passenger Travel Demand Model ...... 50 7.2 Estimating Mode Shares ...... 54 7.3 Example Mode Share Calculation ...... 59 7.4 Scenario Analysis ...... 61 7.4.1 Scenario 1: New Route between Grand Forks and Minot ...... 61 7.4.2 Scenario 2: New Route between Bismarck and Williston via Dickinson ...... 64 7.4.3 Scenario 3: Add a Stop in Wahpeton ...... 66 7.4.4 Scenario 4: Express Service Between Fargo and Bismarck ...... 68 7.4.5 Scenario 5: Gasoline Prices Increase to $5.00 per Gallon ...... 68 8. SUMMARY AND CONCLUSIONS ...... 71 8.1 Implications ...... 72 8.2 Limitations and Further Research ...... 73 REFERENCES ...... 75 APPENDIX A. STATED CHOICE EXPERIMENT DESIGN ...... 81 APPENDIX B. SURVEY ...... 86 APPENDIX C. SAS CODE ...... 95 APPENDIX D. SAS PRINTOUTS ...... 100 Nested Logit Results ...... 100 Multinomial Logit Results ...... 104 Mixed Multinomial Logit Results ...... 108 Binary Logit Results: Automobile Choice ...... 112 Binary Logit Results: Air Choice ...... 117 Binary Logit Results: Bus Choice ...... 122 Binary Logit Results: Rail Choice ...... 127 APPENDIX E. OBTAINING DEMOGRAPHIC DATA IN GIS ...... 132 LIST OF TABLES

Table 3.1 Factors, Modes, and Levels for Stated Preference Questions ...... 13 Table 3.2. Price Levels by Mode ...... 14 Table 3.3. Travel Time by Air ...... 14 Table 3.4. Access and Egress Distance for Air, Bus, and Train ...... 14 Table 3.5. Stated Choice Experiment Factors and Levels ...... 16 Table 3.6 Age Distribution of Survey Sample and North Dakota Population ...... 19 Table 4.1 Survey Response Rate...... 25 Table 4.2 Demographic Characteristics of Survey Respondents and Target Population ...... 26 Table 4.3 Geographic Classification of North Dakota Areas ...... 28 Table 4.4 Response to Opinion Questions ...... 32 Table 4.5 Differences in Traveler Attitudes from 2009 to 2015 ...... 33 Table 4.6 Mode Choice Data from the Stated Preference Survey by Individual Characteristics ...... 35 Table 4.7 Age Composition by Mode Chosen in Stated Preference Survey ...... 35 Table 4.8 Mode Choice Data from the Stated Preference Survey by Trip Characteristics ...... 36 Table 4.9 Mode Choice Data from the Stated Preference Survey by Mode Characteristics ...... 37 Table 6.1 Estimated Mixed Logit Mode Choice Model ...... 45 Table 6.2 Impacts of Attitudes on Mode Choice, Results from Binary Logit Model ...... 49 Table 7.1 Travel Time between Intercity Bus Stops ...... 53 Table 7.2 Travel Time between Intercity Rail Stations ...... 53 Table 7.3 Estimated Bus Mode Shares for Personal Trips...... 57 Table 7.4 Estimated Bus Mode Shares for Business Trips ...... 58 Table 7.5 Data for Personal Trips Originating in TAZ 33 and Ending in TAZ 41 ...... 60 Table 7.6 Estimated Bus Mode Shares for Personal Trips, Scenario 1 ...... 63 Table 7.7 Estimated Bus Mode Shares for Personal Trips, Scenario 2 ...... 65 Table 7.8 Estimated Bus Mode Shares for Personal Trips, Scenario 3 ...... 67 Table 7.9 Estimated Bus Mode Shares for Personal Trips, Scenario 4 ...... 69 Table 7.10 Estimated Bus Mode Shares for Personal Trips, Scenario 5 ...... 70 Table A.1 Coded Survey Design ...... 81 Table A.2 Uncoded Survey Design...... 83 LIST OF FIGURES

Figure 3.1 Two-Level Nest Structure ...... 22 Figure 4.1 North Dakota Regions based on Three-Digit Zip Codes ...... 27 Figure 4.2. Transportation Characteristics of Survey Respondents ...... 29 Figure 4.3. Survey Responses on How Easy it is to Travel to Other Cities in North Dakota ...... 30 Figure 4.4. Survey Responses on Need for Intercity Transportation Improvements in North Dakota ...... 31 Figure 4.5 Mode Choice Responses from Stated Preference Survey ...... 34 Figure 7.1 Traffic Analysis Zones for North Dakota Travel Demand Model ...... 50 Figure 7.2 North Dakota Intercity Bus Network, February 2016 ...... 51 Figure 7.3 North Dakota Intercity Rail Network ...... 51 Figure 7.4 TAZs within 25 Miles of Intercity Bus Stop ...... 55 Figure 7.5 TAZs within 25 Miles of Intercity Rail Station ...... 55 Figure 7.6 Most Populated Cities in North Dakota and Corresponding TAZs ...... 56 Figure 7.7 TAZs Highlighted for Example Calculation...... 59 Figure 7.8 New Service to Dickinson and Watford City Provided in Scenario 2 ...... 64 Figure 7.9 New Service to Wahpeton Provided in Scenario 3 ...... 66 Figure E.1 North Dakota Population Aged 18-24, by Census Block Group ...... 134 Figure E.2 North Dakota Households with Income less than $25,000, by Census Block Groups ...... 134 Acknowledgments

The funds for this study were provided by the United States Department of Transportation through the National Center for Transit Research within the Center for Urban Transportation Research at the University of South Florida. The Small Urban & Rural Transit Center within the Upper Great Plains Transportation Institute at North Dakota State University conducted the research.

This research was conducted and written as a dissertation in fulfillment of the requirements for a Ph.D. in Transportation and Logistics at North Dakota State University. The author gratefully acknowledges the support of his dissertation committee: Amiy Varma, Jill Hough, Canan Bilen-Green, Rhonda Magel, and Fred Fravel. ABSTRACT

The general objective of this research was to develop an intercity mode choice model that could be incorporated into a statewide travel demand model to estimate demand for rural intercity bus services. Mode choice models predict the likelihood of an individual choosing a given mode for a given trip based on individual, mode, and trip characteristics, using a discrete choice modeling technique. These models are used to estimate mode splits and can be used to predict mode volumes when used with known or estimated total trip volume.

Four intercity transportation modes were considered in the study: automobile, bus, rail, and air. Specific objectives were to: estimate the impacts of mode, trip, and individual characteristics on mode choice; estimate the impacts of attitudes on mode choice; examine changes in attitudes regarding intercity travel; and conduct scenario analyses to demonstrate how the model could be used to estimate the possible impacts of service changes or changes in other variables.

The model was applied to the state of North Dakota. To obtain the data needed for the mode choice model, a stated preference survey was conducted of individuals across the state. Then, a mixed logit model was used to estimate the mode share model. Survey

The survey included four sections: a general demographic information section, a section on transportation experience, a stated preference section, and a section on travel attitudes. The stated preference section of the survey presented a series of choice sets to survey participants, where they were asked to identify their mode of choice given varying levels of price, travel time, access distance, egress distance, and frequency of service for each mode, while also considering trip distance, trip purpose, and whether they were traveling alone or in a group.

The target population for the survey was adults aged 18 or older living in North Dakota. Surveys were distributed to a random sample of 5,000 residents. Half of these residents received a paper survey in the mail, and the other half received a postcard by mail with a web address for where they could complete the survey online. A total of 541 responses were received. The paper survey had a 17% response rate, while the online-only survey had a 6% response rate, yielding an overall response rate of 11%.

With each survey respondent given nine different stated preference questions to answer, there were a total of 4,724 stated preference responses received. The automobile was the mode of choice in 73% of these responses, while air, rail, and bus accounted for 13%, 10%, and 4% of responses, respectively. The raw data showed differences in mode shares based on gender, age, income, trip distance, trip purpose, party size, travel costs, travel speeds, and access distances. Mode Choice Model

A mixed logit model was developed to estimate the impacts of individual, trip, and mode characteristics on choice of mode. Among the individual characteristics, gender, age, income, and disability were all found to have some effect on mode choice. Men were found to have greater odds of choosing automobile or bus travel. Older adults aged 70 or older were found to be less likely to choose air travel, and younger adults, aged 18-24, were found to be more likely to choose bus travel. As income increased, individuals were more likely to choose automobile and air travel. The impact of having a disability on mode choice was also found to be statistically significant and large in magnitude. People with a disability were much less likely to choose automobile or air travel and, therefore, more likely to choose bus and rail travel, compared to those without a disability. The results showed that party size and trip purpose have some significant effects on mode choice. Individuals traveling alone were found to be less likely to choose automobile travel, and those traveling for business purposes were more likely to choose air and less likely to choose automobile. Travel time and travel price (consisting of fares and price of gasoline) were found to have negative and statistically significant impacts on mode choice, as did access distances for bus, rail, and air travel. The impacts of egress distance and service frequency, however, were not found to be statistically significant. Finally, all other variables equal, respondents were significantly more likely to choose automobile travel (in comparison to all other modes) and air travel (in comparison to bus and rail), and they were least likely to choose bus. Traveler Attitudes

Comparing responses to the survey’s attitudinal questions to those from a similar survey conducted in 2009 showed little change in traveler attitudes. Survey responses showed a demand for predictability, comfort, and cleanliness. Modeling the impacts of traveler attitudes on mode choice revealed that attitudes play a role. For example, individuals more likely to choose bus than the automobile included those who do not mind traveling with strangers, those more concerned about getting into an accident, those more concerned about having a stress-free trip than reaching their destination quickly, those more sensitive to cost, and those who feel that people who use intercity bus are like them. On the other hand, people who value a more predictable travel time or who place a higher value on cleanliness were more likely to choose the automobile. Implications

The study demonstrated how the model could be used to estimate mode shares between origin-destination pairs after calculating in-vehicle travel times for each mode and access and egress distances to intercity bus and rail stations. When combined with travel volume estimates from a statewide travel demand model, this model would provide estimates for travel volume by intercity bus or rail.

The model could be used to estimate changes in mode shares and travel volume by each mode under different scenarios, including changes in service characteristics, demographic characteristics, or costs of competing modes. The model could be used to estimate demand for a new route, and the feasibility of the route could be analyzed by comparing the estimated ridership and fare revenue with the expected costs, allowing for a cost-benefit analysis.

Similarly, the model could be used to estimate the impact of adding a new stop along an existing route. The new stop would provide access to more potential users, but it would also increase travel time along the route. On the other hand, a low-usage stop could be removed to decrease travel time along a route. This model could be used to evaluate the tradeoff between providing access to more potential users or decreasing travel time. Two of the scenarios analyzed in this study evaluated such a tradeoff.

The model also takes into account how individual routes fit into a larger network. Ridership on an individual route can affect or be affected by ridership on another route. The model would be able to determine the overall net effect on ridership from the addition of the route, and its feasibility could be assessed. Policy makers would be able to use the tool to determine if the additional ridership could justify public investment.

In addition to changes in service characteristics, the model could be used to analyze impacts of demographic changes or changes in the characteristics of competing modes. Notably, the price of gasoline can have a significant impact on the demand for intercity bus service. One of the scenarios in this study analyzed the impact of higher gasoline prices. Bus mode shares approximately doubled when the price of gasoline increased from $2.00 per gallon to $5.00 per gallon, although this scenario did not consider a corresponding increase in bus fares.

With an aging population in North Dakota and across the county, consideration of demographics is important. The results showed that adults aged 70 or older are much less likely to travel by air, and given that driving rates tend to decrease for older adults, demand for intercity bus could increase as the elderly population continues to increase. The results did not specifically show that older adults are more likely to choose intercity bus, compared to those who are younger, but it did show that people with disabilities are much more likely to choose bus or rail. Given that disability rates increase significantly with age, an aging population would lead to increased demand for intercity bus or rail services.

The analysis of attitudes and mode choice provides some insight into how intercity bus services are perceived by the public and how they could be marketed. Bus services tend to be perceived as less stressful than automobile or air services and safer than the automobile. Intercity bus companies could market their services to those more concerned about having a safe and stress-free trip, those who want to make a more productive use of their time when traveling, or those who like to visit with others when traveling. Results also suggest intercity bus service is perceived as less predictable and less clean than traveling by automobile, whether warranted or not, and that individuals would be more likely to travel by bus if they were made to feel like they are similar to other bus users. 1. INTRODUCTION

A significant need exists for a model to estimate demand for intercity bus services, especially in rural areas. Many states and rural operators are unsure about the potential demand for rural intercity bus service, and many existing models are unreliable due to poor data (Fravel et al. 2011). Numerous types of intercity demand models have been estimated over the past few decades, but these models have their limitations, and intercity modeling remains less developed than urban travel demand modeling.

To address the need for rural intercity bus demand modeling, a TCRP project by Fravel et al. (2011) developed a sketch-planning guide that could be used by state transportation department program managers and both public and private rural intercity bus service providers to forecast demand for rural intercity bus services. The route-level modeling techniques used in this TCRP report provide a useful tool for estimating ridership on rural intercity routes, but it has some limitations. It does not account for through passengers using the service, and it is not sensitive to changes in fares or frequency.

Many previous demand models are route-level, corridor-level, or city-pair models. While these models can be useful, they ignore the effects of existing within a larger network, and they rely on aggregate data. Disaggregate data, or data at the level of the individual or household, can be more useful in developing travel demand models (Koppelman and Hirsch 1986). Intercity mode choice models have been developed using disaggregate data, but there are many variables that could influence mode choice which are often not included in these models.

Demand for intercity transit services can be estimated within the framework of a statewide intercity travel demand model that has a mode choice component. A number of states have operational statewide travel demand models, but they are often rudimentary or lacking in necessary detail. Many are partial models or focus on freight. Miller (2004) argued that many passenger models do not pay sufficient attention to access and egress when estimating mode choice splits. Access and egress refer to the distance to and from the long-distance mode at either end of the trip.

Previous research by Mattson et al. (2010a,b) estimated a mode choice model for regional intercity travel for residents of North Dakota and western Minnesota. The study used a stated preference survey and a multinomial logit model to estimate the impacts of mode characteristics, including travel time, cost, and frequency, and individual characteristics, including age, income, etc., as well as trip characteristics, including distance and trip purpose, on choice of mode. The model did not consider issues of access and egress but assumed respondents had easy access to each of the modes. In reality, of course, many travelers in rural areas need to travel considerable distance to reach a transit station. Information is needed on how far travelers are willing to drive to a bus station. Adding stops along a route would increase the number of potential users with access to the service, but it could also have a negative impact on ridership as it increases travel time on the route. The impacts of both of these factors need to be considered.

1.1 Research Objectives

The general objective of this research is to develop an intercity mode choice model that can be incorporated into a statewide travel demand model to estimate demand for rural intercity bus services. Four intercity transportation modes are considered in the study: automobile, bus, rail, and air. Specific objectives are to:

 Estimate impacts of mode characteristics, such as cost, in-vehicle travel time, access and egress times, and service frequency on intercity mode shares.

1  Estimate impacts of individual characteristics, such as age, income, gender, and disability on intercity mode shares.  Estimate impacts of trip characteristics, such as travel distance, trip purpose, and party size on intercity mode shares.  Estimate impacts of attitudes on intercity mode shares.  Examine changes in attitudes regarding intercity travel.  Develop a mode choice model that can be incorporated into a North Dakota statewide travel demand model to estimate demand for intercity bus services.  Conduct scenario analysis to estimate the effects of possible service changes or other changes.

This study develops a mode choice model that can be used to estimate the percentage of trips used by each mode for each origin-destination (O-D) pair throughout the state of North Dakota given the current intercity transit network and hypothetical changes to the network. This model is developed so that it can be incorporated into a statewide travel demand model under development. The number of trips by each mode between each O-D pair could then be estimated using results from the mode choice model, with data for travel time by each mode and access and egress times for rail and bus.

Mode shares are first estimated using current conditions, including demographic data, gasoline prices, and service conditions. Then additional scenarios are run under different mode characteristics. The model can be used to estimate the demand for potential new transit services, new transit stops, reduced travel time, or other service changes, as well as impacts from demographic changes or changes in competing modes such as rising gasoline prices.

1.2 Organization

Previous research on intercity bus demand modeling is presented in Section 2. This includes a discussion of basic and modified gravity models; route-level, corridor-level, and city-pair models; mode choice models; statewide travel demand models; and other approaches. This study employs a mode choice model estimated with a mixed logit model using data collected from a statewide stated preference survey. The methodology and development of the survey are detailed in Section 3. Survey results are presented in Section 4. In Section 5, the final mode choice model specification is described. Results are provided in Section 6. Section 7 details how the mode choice model can be incorporated into a statewide travel demand model, and shows estimated mode shares for intercity bus within the state. This section also includes a series of alternative scenarios showing how mode shares would change under different conditions or service characteristics. The final section provides a summary, a discussion of implications and limitations, and conclusions.

2. LITERATURE REVIEW

Most previous research on intercity bus demand consists of corridor or route studies and mode choice studies. Researchers began developing intercity corridor models in the 1960s, motivated largely by a need to evaluate investment alternatives for improving travel in the northeast corridor (Washington-New York- Boston). These models have evolved, and different approaches for studying intercity passenger travel have emerged, but these models lack the complexity of urban travel demand models (Miller 2004). 2.1 Basic Structure of Existing Models

As Fravel et al. (2011) explained, intercity demand studies often start with the basic gravity model, which generally states that demand for travel between two places is proportional to the populations and inversely proportional to the distance between them. This relationship can generally be formulated as the population of one urban area times the population of the second, divided by the distance between them squared. The simplest gravity models use city population to measure attractiveness and distance as the measure of impedance. Some basic gravity models also include income. Modified gravity models include measures of impedance such as travel time, travel cost, service frequency, and generalized cost of travel, as well as a more extensive list of city attractiveness variables.

2.1.1 Major Ridership Generators

Studies of intercity bus demand need to go beyond the simple gravity model to include places that are potential attractors of intercity bus ridership. Previous studies have included colleges and universities, major military bases, hospitals and major medical facilities, regional correctional facilities, recreation areas, and major intermodal connections at airports as facilities that would attract ridership (Utah Department of Transportation 2010, KFH Group 2003, Yang and Cherry 2012). Yao and Morikawa (2005) calculated attractiveness using factors such as number of headquarters, number of international conferences, and other business-relevant factors for business trips. For non-business travel, they used factors capturing suitability for leisure activities, such as number of resorts, sport centers, museums, cinemas, shopping centers, etc.

2.1.2 Characteristics of Population

The demand model can be revised to account for both total population served, as well as the characteristics of that population. Studies have shown that certain populations have a greater likelihood of using intercity bus services (Yang and Cherry 2012; KFH Group 2001, 2003; Fravel et al. 2011, Al-Sahili and Sadeq 2003, Kack et al. 2011, Sperry et al. 2014). These populations may include older adults, youth (aged 18-24), low-income persons, people with mobility limitations, and people without access to an automobile. In a survey of intercity bus riders in Montana, Kack et al. (2011) found the 18-24 age group most likely to use the bus, and half of all respondents were from a household with income of less than $15,000 per year. Sperry et al. (2014) found similar results in a survey of intercity bus riders in Michigan. 2.2 Route, Corridor, and City-Pair Models

Many previous studies estimate ridership along a route or corridor or between two cities. Fravel et al. (2011) discussed demand models for boardings at a stop, route-level demand models, and city-pair demand models. A demand model for boardings at each stop, or a point demand model, could be estimated as a function of population and demographic data for the bus stop location and characteristics of the service. This demand model uses demographic and service data to estimate the number of persons boarding at a particular stop, similar to models estimated by KFH Group (2001, 2003). Such a model

3 would require data on boardings or ticket sales from intercity bus carriers, service data (e.g., frequency, fares) from the intercity bus carriers, and population and demographic data from the Census.

A route-level demand model is a similar model that could be used when there are several stops along a route that generate ridership. Calibrating such a model would require ridership for the entire route or route segments, fare and frequency data, and demographic data. Both the point demand model and the route- level demand model lack the ability to estimate potential overhead traffic. A city-pair demand model can be used to estimate ridership between a particular pair of cities. Many previous studies, including most of those reviewed by Koppelman et al. (1984), used city-pair demand models.

2.2.1 Overview of Early Research Efforts

Koppelman et al. (1984) reviewed the development of intercity passenger demand models from the 1960s through the early 1980s. Many of these models were corridor models using aggregate travel data. They described five groups of aggregate corridor models developed during this period: direct origin-destination (O-D) mode volume models, total O-D volume (all modes) models, mode share models, sequential models of total intercity demand and mode share, and direct demand models for a single intercity mode.

The direct O-D mode volume models predict travel demand for each O-D pair separately for each mode using one or a set of equations. These models use variables to describe the city pair, including population, per capita income, and employment, and variables to describe the modes, including travel time, travel cost, and service frequency. According to Koppelman et al. (1984), these models made a number of contributions but contain a number of weaknesses, such as the exclusion of a number of important travel service variables (e.g., comfort, convenience, reliability, safety), inadequate representation of population distribution characteristics (e.g., income segmentation), lack of access times and costs to intercity terminals, no clear basis to define market area for the city or terminal, and other failings.

The total O-D volume models estimate total demand between city pairs using a single equation with a gravity-type formulation. These models do not predict mode volumes or mode splits, and they typically do not include mode characteristics or important policy variables.

Mode share models predict mode splits and can be used to predict mode volumes when combined with total intercity travel volume. Sequential models of total intercity demand and mode share are two-stage models that combine the total volume models with mode share models. These are similar to the sequential model system used in urban travel demand modeling, and the main advantage of such an approach is that it estimates the effects of policy changes separately on city-pair trip volume and intermodal competition (Koppelman et al. 1984).

Some direct-demand, non-sequential models estimate demand for a single mode. These generally are formulated using a simple or modified gravity model. Population, income, and distance between cities are used in the most basic gravity models. While modified gravity models can estimate impacts of policy changes and are an improvement over simple gravity models, the models reviewed by Koppelman et al. (1984) do not account for intermodal competition.

Some time-series analysis has also been conducted to capture time trends, seasonal variations, period-to- period correlations, and lagged response in ridership in response to changes in service characteristics or other variables. These single-mode models, according to Koppelman et al. (1984), are designed to be used in cases where there is a particular interest in only one of the modes in a corridor. One advantage to using such a model is the reduced requirements for data, but these models suffer from a lack of intermodal competition in the formulation.

2.2.2 Route-Level Research for Rural Intercity Bus Service

Burkhardt and Riese (1982) developed a model to estimate ridership on rural intercity bus routes as a function of route length, the frequency of service, and the population served. The population data used in the model included the total population of each city, town, or village lying adjacent to the route, or adjacent to a town served. Population data for the central city or town center served were used, rather than for the entire metropolitan area. The model used least squares regression and was developed using data from Greyhound for 89 routes in 17 states. As KFH Group (2001) noted, this research was published in 1982, and estimates from a model developed this long ago may no longer be relevant. The overall level of intercity bus ridership has changed over time, so estimates from old data would not be reliable.

Because of the age of this model, KFH Group (2001) developed a new model to estimate demand for a new route between Macon, GA, and Brunswick, GA. They developed a linear regression model to estimate passenger revenue as a function of the attributes of the location. To estimate the model they obtained passenger revenue and passengers by stop for all Georgia locations for 1999 and 2000, and used data for 18 small towns similar to the ones along the proposed route. They hypothesized that the following would affect intercity bus revenue: the 10- and 20-mile market area population; the percentage of the population with incomes below the poverty line; the percentage of households with no vehicle; the percentage of the population with incomes under $25,000; the frequency of bus service, measured in departures per week; and a dummy variable for Greyhound service within 10 miles. However, only total population and frequency were found to be statistically significant.

KFH Group (2003) followed this study with a similar analysis to examine the feasibility of implementing intercity bus service between the cities of Hampton, VA, and Fredericksburg, VA. To do so, they developed a model using data from Greyhound for passenger revenue for all Greyhound stops in eastern North Carolina, Virginia, and Maryland. A linear regression model was developed that estimated revenue as a function of attributes of the location. Data were restricted to stops in towns with a population of less than 20,000, so that it would be more useful for predicting revenue to the small towns along the proposed route.

The following factors were hypothesized to affect intercity bus ridership: the population within the town's boundaries, the 10-mile population around the stop, the percentage of the population with income below the poverty level, the percentage of households that rent, the percentage of the population older than age 60, the percentage of the population aged 18-24, the frequency of bus service, measured in departures per week, and the presence of a four-year residential college, major medical facility (more than 150 beds), military installation, state or federal level correctional institution, or a locally operated transit system. Dummy variables were included for the presence of these special generators.

Results showed that population within 10 miles of the stop, the percentage of population below the poverty line, the presence of a medical facility, and service frequency all had a significant positive effect on annual revenue. A drawback of this model is that it only estimates revenue generated en-route, and it does not include any "overhead" traffic revenue, or traffic generated by those traveling beyond these points.

More recently, Fravel et al. (2011) developed a route-level model based on the theory that rural intercity demand is a function of the following: overall population levels of origin points, population of the destination city, population characteristics, length of the route or service, basic service characteristics (frequency, fare level, etc.), impact of key institutions that are likely to concentrate demand, and connectivity of the service. These factors were used to build upon the basic gravity model, where the friction factor includes actual distance, the fare level, and the frequency. This work was published as part of a TCRP project described in TCRP Report 147.

Fravel et al. (2011) developed a regression model to predict ridership as a function of the populations served and the service characteristics. They found that using corridor population alone provides poor estimates. Separate models were estimated for regional providers and intercity bus providers. Population was divided into urbanized and non-urbanized. The final model estimated ridership on a route (annual one-way passenger boardings) as a function of origin population (sum of populations of origin points along the route, excluding the point with the largest population which is referred to as the destination population), the number of stops on the route, whether it provides a connection to an airport, and whether service is operated by a carrier meeting the definition of an intercity carrier. This model provided much greater predictive power, and it showed that ridership along the route increased with an increase in origin population, number of stops on the route, and the presence of an airport connection. Ridership was also found to be greater for providers defined as intercity carriers. The impact of the airport connection and intercity bus carrier demonstrate the advantages of connectivity. An intercity bus provider, as opposed to a regional transit provider, provides the advantages of interlining, allowing greater connectivity to the intercity bus network.

Fravel et al. (2011) also examined long-distance trip rate data from the 2001 National Household Travel Survey (NHTS). They obtained trip rates categorized by urbanized and non-urbanized, region, and income. A problem with using the NHTS data is too few trips reported by intercity bus. Therefore, they used overall trip rates for long-distance travel and applied a mode split factor, which they assumed to be 0.09% for bus. This trip rate method, however, was found to be less accurate than the regression model. 2.2.3 Other Studies

Al-Sahili and Sadeq (2003) studied intercity bus ridership in Palestine using ridership data obtained from bus operators there. Their model predicted ridership between city pairs, and five independent variables were examined: origin city population, destination city population, bus fare, percentage of population in origin city consisting of students attending secondary schools or universities, and percentage of population in origin city consisting of people older than 15 years who are employed. The study found that origin and destination city population, percentages of employees and students, and bus fare all significantly influenced ridership between city pairs.

Wirasinghe and Kumarage (1998) developed an intercity demand model for use in a developing country. The model form was calibrated for Sri Lanka, but can be applied to other countries. Their model formulation for total trips between districts is a modified gravity model, including socioeconomic variables and impedance variables. Impedance variables include wait time, transfers, service frequency, and generalized cost of travel (which is a composite cost variable of minimum travel time and minimum fare). Wirasinghe and Kumarage (1998) used the following socioeconomic variables in developing a model appropriate for Sri Lanka: the product of origin and destination populations, the product of urban population proportions (based on the assumption that urban areas generate more travel than rural areas), and the difference of urban population proportions (based on the assumption that there would be less need to seek services outside the district in districts with a large urban center).

Pagano et al. (2001, 2003) estimated demand between city pairs in Illinois using a gravity-type formula. Their rationale was that the level of demand for intercity bus service between two places, excluding intermediate stops, would be proportional to the population masses of the terminal places and inversely proportional to the squared distance between places. For their study, they did not have data on the number of trips between city pairs, so they developed a scoring function to determine the interactivity potential between city pairs, which was equal to the product of the populations divided by squared distance. The method then assigned a score to each O-D pair and sorted the pairs in descending order. This method is useful for identifying the greatest need for service, but it does not provide ridership estimates and is limited in that it only accounts for population and distance.

Yao and Morikawa (2005) developed an intercity travel demand model to account for induced travel. As they note, ridership is determined by both exogenous factors that determine the location of the demand curve, such as population and economic development, and endogenous factors that determine the point along the demand curve, such as changes in travel time and travel cost. Reducing travel time and cost will lead to increased travel, and this induced travel may include shifts between alternative routes, modes, destinations, times of day, or new trips not previously made (Yao and Morikawa 2005). To capture induced travel, Yao and Morikawa (2005) developed an integrated model including trip generation, destination choice, mode choice, and route choice. As they noted, the destination choice utility function is generally specified using variables to describe the city’s characteristics and a measure of impedance between cities. 2.3 Intercity Bus Network Model

Fravel et al. (2011) proposed a network model that would follow the structure of the urban travel demand model and apply it to the intercity bus network. They described a single-mode bus model that would include trip generation and trip distribution to a network but would not include mode split. As they described, such a network could be used by bus companies and state DOTs to evaluate network impacts of adding links to the network or bypassing congested stops, and a national intercity bus network model could be developed which could estimate impacts from service changes, such as rerouting service, adding new links, or eliminating routes or links. The network model would describe the initial network in terms of the routes, bus stops, frequencies, and travel times between stops.

In this framework, a point demand model would be used to estimate the number of intercity bus trips generated at each stop. The point demand model would be used for the trip generation step. Once the number of boardings at each stop has been estimated, trip distribution and assignment would place the trips onto the links, resulting in the overall ridership on each link. A type of gravity model could be used for calculating trip distribution, and O-D data or data on the distribution of trip distances for intercity bus trips would be needed to calibrate the model.

There are advantages to analyzing the intercity bus network from a statewide or regional level rather than at the corridor level. Because each individual route operates within a larger network, its performance is likely to depend on the other parts of the network. Because of the substitutive and complementary effects among intercity bus routes, Guo et al. (2008) argued that analyses that focus only on the individual corridor would lead to erroneous conclusions. 2.4 Disaggregate vs. Aggregate Models

The models previously described are aggregate models that rely on total ridership numbers. Disaggregate models differ from aggregate models in that they use data at the level of the individual or household. These models are used widely in urban travel analysis, and according to Koppelman et al. (1984), they provide better results and are preferred because of their ability to represent the individual traveler’s decision process. These models require data describing the travel behavior of a set of individuals, their individual characteristics, and the attributes of the travel services available to each of them.

As described by Koppelman et al. (1984) the use of aggregate data results in parameter estimation bias. Other weaknesses they describe are that the formulation of the models is correlative rather than causal, the selection of variables is generally based on statistical results and may not represent the true determinants, limitations on data availability and measurement problems may undermine the quality of the models, the models are not able to consider the relationship between intercity travel decisions and related travel decisions, and the validation of the models has been limited, partially due to lack of data.

Because of the problems with existing models using aggregate data, Koppelman and Hirsh (1986) recommended a disaggregate approach. Their study presented a conceptual structure of the intercity passenger decision process. They recommended using a fully disaggregate dataset, with data gathered at the individual or household level, compatible with a behavioral framework. The recommended dataset would include individual characteristics, actual intercity travel behavior over a substantial period of time, full description of all of the intercity trips undertaken during this period (purpose, party size, time of year), relevant information about the destinations visited (city, specific areas visited, number of stops, trip duration), attributes of the modes chosen for the trips as well as the corresponding attributes of the non- chosen modes, and description of the local activity pattern at the destination (length of stay, accommodations, and transportation arrangements). Koppelman and Hirsh (1986) noted that no existing datasets included all the desired information.

Most of the early research reviewed by Koppelman et al. (1984) used aggregate data, but they also described a few disaggregate intercity travel models that included binary mode choice, multinomial mode choice, and destination-mode choice models. More recently, a number of disaggregate intercity mode choice models have been developed. 2.5 Mode Choice Models

Mode choice models predict the likelihood of an individual choosing a given mode for a given trip based on individual, mode, and trip characteristics, using a discrete choice modeling technique. These models are used to estimate mode splits and can be used to predict mode volumes when used with known or estimated total trip volume.

Factors affecting mode choice can be organized into three categories: the characteristics of the different transportation modes, the characteristics of the individual making the trip, and the characteristics of the trip itself. Mode characteristics that could potentially affect mode choice include cost, travel time, comfort and convenience, service frequency, need for transfer, and access. The trip-maker characteristics include income, age, gender, car-ownership, ability to drive, and preferences and attitudes. Trip characteristics include trip purpose, trip length, and size of party.

Most studies find that cost and time are important determinants of intercity mode choice. Cost is commonly regarded as one of the main factors, and a variety of intercity and mode choice studies over the last several years have demonstrated the importance of travel time (Kumar et al. 2004, Ashiabor et al. 2007, Proussaloglou et al. 2007, Andrade et al. 2006). Studies also include other mode characteristics such as service frequency. Access, egress, and transfer times and costs should also be accounted for to compare door-to-door travel costs (Zhang et al. 2012).

Most studies include some socioeconomic variables as predictors of mode choice. Many include income (Kumar et al. 2004, Ashiabor et al. 2007, Proussaloglou et al. 2007), while age, gender, education, and profession have also been considered (Kumar et al. 2004). Socioeconomic factors can affect how sensitive travelers are to travel time and cost. For example, some research has shown that high-income travelers are less sensitive to travel cost (Ashiabor et al. 2007). The habits and attitudes of individuals can also be important. A number of studies have investigated the effects of habit on mode choice (Aarts et al. 1997, Verplanken et al. 1997, Garvill et al. 2003), while others have asked respondents a series of attitudinal questions and analyzed how different attitudes influenced choice of mode (Gilbert and Foerster 1977, Golob and Recker 1977, Outwater et al. 2003, Mattson et al. 2010b, Ripplinger et al. 2011).

Finally, the characteristics of the trip itself can influence the choice of mode. Research has shown that transit is a closer substitute to the automobile for commuter or business trips than for leisure travel (Storchmann 2001). Business travelers may be motivated differently than those traveling for personal reasons. Also, as the trip distance increases, the substitutability between the different modes may also

8 change. For example, motorists may be more likely to switch to bus or rail in response to higher gas prices as the length of the trip increases (Wallis and Schmidt 2003, Currie and Phung 2008), which could be due to an increase in the cost difference at greater trip distances. The size of the travel party could also be an important variable that is commonly ignored in mode choice studies (Miller 2004). As the size of the travel party increases, the automobile becomes more cost-effective.

Mode choice models are often estimated using data collected from stated preference (SP) surveys. SP surveys, also referred to as stated choice experiments, are widely used in areas such as marketing and transportation. A number of transportation studies have utilized an SP survey to analyze transportation alternatives (Mattson et al. 2010a,b, Kumar et al. 2004, Andrade et al. 2006, Dehghani et al. 2002, Mehndiratta and Hansen 1997, Pinjari and Bhat 2006, Richardson 2002, Richardson 2006). These studies analyze the SP survey data using a discrete choice modeling technique. Discrete choice modeling is popular in transportation and marketing research for understanding an individual’s stated choice among alternatives (Kuhfeld 2010). The multinomial logit model, which is a type of random utility choice model, has been traditionally used to model the choice among alternative modes in intercity travel demand modeling (Koppelman and Sethi 2005). The nested logit model has been shown to be superior to the traditional multinomial logit model for intercity model choice studies (Forinash and Koppelman 1993), as has the mixed logit model (Hensher and Greene 2003). Use of these more advanced models has increased in recent years. 2.6 Statewide Travel Demand Models

Intercity mode choice can be included as part of a statewide travel demand model. Some states have travel demand models that include intercity bus as one of the modes. Statewide travel demand models follow the same basic four-step process as the more common urban travel demand model and are run using GIS. A statewide travel demand model divides the state into a number of traffic analysis zones (TAZs). Population, employment, and other characteristics of each TAZ are used to estimate the number of trips produced and attracted by each zone. Trips between zones are distributed using a gravity model and assigned to the transportation network based on minimizing travel times. Some of these models also include the intercity bus network and a mode choice component that assigns a certain percentage of the trips to the bus network.

The direct demand model is an alternative to the four-step process. In a typical direct demand model, the aggregate passenger travel demand between an O-D pair by each transportation mode is expressed as a function of economic, land use, and socio-economic characteristics of the origin and destination, as well as the attributes of the particular mode and its competing modes (e.g., travel time, cost, other level of service factors) (Zhang et al. 2012).

Statewide travel demand models, though, have typically been motivated by freight concerns and model truck traffic (Giaimo and Schiffer 2005). These models are not as advanced as urban models in terms of data collection and modeling (Zhang et al. 2012). Not all states have statewide travel demand models, and many of the existing models have been developed within the last couple of decades. One of the biggest problems with these models is a lack of data on how and why U.S. travelers choose different long- distance modes, destinations, and routes (Zhang et al. 2012).

A NCHRP project, NCHRP Synthesis 358 conducted by Horowitz (2006), described the state of statewide travel demand models. Horowitz (2006) noted that the state of practice had been maturing over the previous 10 years and, at the time of his publication, approximately half of the states had functional models. Most states tend to rely on secondary data sources, such as the National Household Travel Survey or the Census Transportation Planning Package.

One of the more advanced statewide travel demand models that incorporates the intercity transit network is one developed for the Wisconsin Department of Transportation (WisDOT) by Proussaloglou et al. (2007). Their passenger model was estimated using Wisconsin add-on data for the 2001 National Household Travel Survey (NHTS), and model validation relied on the NHTS data, statewide traffic counts, intercity transit ridership estimates, and 2000 U.S. Census data including the Census Transportation Planning Package. The model divided the state into 1,642 zones, derived the highway network from WisDOT databases, and developed the intercity transit network to reflect all the existing intercity rail and bus services serving Wisconsin. Their mode choice model included in-vehicle travel times, access and egress travel times, walk access to existing intercity services, frequency of service, transit fares, vehicle availability, household income, the size of traveling party, and the need for transfers as factors that would impact choice of mode. 2.7 Other Approaches

Fravel et al. (2011) also described a few different approaches for estimating rural intercity bus demand: per capita intercity trip generation rates, use of comparable services, and use of historical data. The per capita intercity trip generation rate approach estimates the number of trips per year per capita for a given town and multiplies it by population to get an estimate of total annual ridership. The model is developed based on existing ridership data. This is a very basic model that does not consider service characteristics, demographic differences, travel patterns, or other factors that would influence use of intercity bus service. Another approach is to compare the service being studied with similar routes or services and use data from the similar services to develop ridership estimates. There are many factors that can influence ridership, however, so an accurate comparison could be difficult. Using historical data from previous services may be possible in some cases, but as Fravel et al. (2011) noted, there are several significant issues with this approach.

Koppelman et al. (1984) described aggregate regional intercity demand models that estimate total intercity travel volume, usually for a single mode, for all the intercity corridors in a selected region. These models can be used to estimate the impact of service changes, price changes, or regional economic changes on total regional ridership, but they cannot be applied to specific corridors. 2.8 Conclusions

Important determinants of intercity bus demand include population, demographics, location of major trip generators, and service characteristics, such as travel time, service frequency, fares, and transfers. Proper analysis should also consider competition between modes, and to be useful, demand models need to consider issues relevant to decision-makers. Models can be evaluated based on their policy relevance, effectiveness in providing useful forecasts, and ease of implementation (Koppelman et al. 1984).

A modified gravity model with attractiveness and impedance factors can be used to estimate demand between city pairs. Attractiveness variables should include population, demographics, and major trip generators; and measures of impedance should include travel time and level of service. However, the gravity model approach becomes more complicated once a route has numerous intermediate stops that offer alternative destinations (Fravel et al. 2011, Cook and Lawrie 2008). Previous studies have commonly developed gravity-type models to estimate demand for a single mode or all modes between city-pairs, while other studies have focused on mode choice modeling.

The Fravel et al. (2011) model and the previous models used by KFH Group are useful for estimating demand on proposed feeder routes, but one of the major drawbacks of these models is that they only take into account characteristics along the route itself while not factoring in how the route fits into a larger network. As a result, predicted ridership does not account for through traffic, or trips that begin or end

10 elsewhere on the network. The new route may simply be filling a gap in the network. Predicting ridership on the route, therefore, requires not just population and service information for that route but also information about how the route fits into the larger network. Therefore, a network model would be more useful than a route-level model for predicting demand.

Ideally, demand for intercity bus service could be estimated within a statewide travel demand model with a mode choice component. Such a model would benefit from the use of disaggregate data. As Koppelman and Hirsh (1986) argued, a disaggregate model with data gathered at the individual or household level is the preferred approach for estimating demand for intercity bus services. Koppelman and Bhat (2006) also noted that disaggregate models are more likely to be transferable to different points in time and different locations. Previous mode choice models have used data at the level of the individual to estimate mode shares for intercity bus and other modes. However, as Miller (2004) showed, there is a wide set of explanatory variables needed for modeling mode choice, and most studies at the time had used a fairly limited set of variables.

3. METHODS

Applying an intercity mode choice model, with disaggregate data, within a statewide travel demand model provides an attractive approach for estimating demand for intercity bus services. As previously noted, however, not all states have statewide travel demand models, and many existing models have been motivated by freight concerns and are not as advanced as urban models.

This study aims to improve upon previous mode choice models by developing a model focusing on passenger travel that can be incorporated into a statewide travel demand model for the state of North Dakota. Unlike a previous mode choice model developed in North Dakota and Minnesota by Mattson et al. (2010a,b), this model includes access and egress distance, which are important determinants of intercity bus use. The model can be used to determine how far travelers are willing to travel to and from intercity bus stops, and by incorporating it into a statewide travel demand model, it can be used to estimate demand throughout the state. The model can be used to estimate ridership on proposed services or the impacts of potential service changes, changes in fares or travel times, rising gasoline prices, or other factors on the use of intercity bus.

To obtain the data needed for the mode choice model, a stated preference survey was conducted of individuals across North Dakota. Then, a discrete choice model was used to estimate the mode share model.

The model includes four modes: air, automobile, bus, and rail. Choice of mode is hypothesized to be affected by different mode-specific factors, generic trip attributes that do not depend on the mode, and individual characteristics. Mode-specific factors include price (the explicit financial cost), travel time, frequency of service, and access and egress distances. Generic trip attributes include trip type (personal or business) and party type (traveling alone or in a group). Individual characteristics include age, income, gender, disability, and personal attitudes. Information for each of these variables was collected through the survey. 3.1 Survey Development

The first two sections of the survey collected demographic information about the respondents, including age, gender, household income, disability, and whether they have access to an automobile and are able to drive. It is expected that these factors may influence mode choice. The survey also asked respondents if they have ever traveled by intercity bus or rail or by airplane. Research has shown that habit plays a role in mode choice (Garvill et al. 2003, Verplanken et al. 1997, Aarts et al. 1997). The survey also asked for the respondents’ zip code, so that the geographic distribution of survey respondents could be documented.

3.1.2 Stated Preference Survey

The next part of the survey was a stated preference (SP) survey. SP surveys, also referred to as stated choice experiments, are widely used in areas such as marketing and transportation. In such a survey, the respondent is shown a number of choice sets. Each choice set consists of two or more options, or alternatives, that are described by a set of attributes with varying levels. The survey respondent is asked to choose his or her preferred option. In this survey, each choice set described a hypothetical intercity trip,

12 and the respondent was asked to choose between four options for taking the trip: air, automobile, bus, or train.

SP surveys have gained popularity for use in travel behavior research due to their ability to accommodate hypothetical alternatives and identify behavioral responses to choice situations that are not revealed in the market (Hensher 1994 and Kumar et al. 2004). An SP survey is useful for this study because it allows for attributes to be varied to levels not yet observed in the market and for individuals to be surveyed who do not currently have access to each of the modes of travel. Furthermore, real-world collection of all the necessary variables could prove to be difficult. Hess et al. (2007) argued that revealed preference (RP) survey data, which attempts to collect information about actual travel behavior, suffers due to the low quality of data relating to un-chosen alternatives. SP surveys also have their disadvantages, one being that actual behavior does not always match what people say they would do (Murphy et al. 2005). However, because of the advantages, a number of transportation studies have utilized an SP survey to analyze transportation alternatives (Andrade et al. 2006, Dehghani et al. 2002, Haghani et al. 2015, Hess et al. 2007, Jiang and Zhang 2014, Kumar et al. 2004, Mattson et al. 2010a,b, Mehndiratta et al. 1997, Pinjari et al. 2006, Richardson 2002, Richardson 2006, and Srinivasan et al. 2006).

Hensher (1994) identified the following key steps to designing a stated choice experiment: 1) identify the set of attributes, or factors; 2) select the measurement unit for each attribute; 3) specify the number and magnitude of attribute levels; 4) create the statistical design, combining the attribute levels into an experiment; 5) translate the design into a set of questions; 6) select an appropriate estimation procedure; and 7) use the estimated parameters to obtain choice probabilities for each alternative. Each of these steps will be described in the following sections.

3.1.2.1 Mode Alternatives, Attributes, and Levels

The survey included four mode alternatives for respondents to choose among: air, automobile, bus, and train. These alternatives were described by a set of mode-specific attributes, and survey respondents also had to consider a set of generic trip attributes that did not depend on the mode.

Generic trip attributes included trip distance, trip type, and party type. One-way trip distances of 50 miles, 100 miles, 250 miles, and 400 miles were used. Trip type was categorized as either personal or business, and the party type indicated if the individual was traveling alone or with a group of either family and friends or co-workers.

Mode-specific attributes included price (the explicit financial cost, which was the price of gasoline for automobile travel or the fare for other modes), travel time, access distance (travel distance from origin to bus stop, train station, or airport), egress distance (travel distance from bus stop, train station, or airport to final destination), and service frequency. Different levels for each factor were used and varied between the choice sets. The factors and their relative levels are presented in Table 3.1.

Table 3.1 Factors, Modes, and Levels for Stated Preference Questions Factor Modes Levels Price Air, automobile, bus, train Low, medium, high Travel time/speed Air Slow (with transfer), Fast (direct flight) Bus, train Slow, medium, fast Access distance Air, bus, train Short, medium, long Egress distance Air, bus, train Short, medium, long Frequency Air, bus, train Three times per week, once per day, twice per day

Each mode had three relative price levels: low, medium, and high (Table 3.2). Automobile costs included just the price of gasoline at the following levels: $2.00 per gallon, $3.50 per gallon, and $5.00 per gallon. Automobile costs were presented in the survey in dollars per gallon rather than cost per mile or total cost per trip because this is the information travelers would encounter in the real world when making a decision. The range of intercity bus and rail prices were determined by reviewing fares charged by Jefferson Lines and Amtrak for trips in North Dakota. Intercity bus fares tend to vary by distance, day of the week, and proximity to travel date. Fares generally are lower on weekdays and if they are purchased further in advance, and fares tend to be higher on a per mile basis for shorter trips.

Table 3.2 Price Levels by Mode Low Medium High Air $300/trip $600/trip $900/trip Automobile $2.00/gallon $3.50/gallon $5.00/gallon Bus $0.15/mile $0.20/mile $0.25/mile Train $0.15/mile $0.20/mile $0.25/mile

Bus and train modes each had three speed levels: slow, medium, and fast. Slow travel was defined as 45 miles per hour, medium speed was 55 miles per hour, and fast travel was 65 miles per hour. Slower travel occurs when there are multiple stops along the route or buses travel along routes with lower speed limits. Automobile speed was assumed to be 65 miles per hour. Air had two speed levels based on if a transfer was required (slow) or if it was a direct flight (fast). Table 3.3 presents travel time by air as a function of trip distance.

Table 3.3 Travel Time by Air Miles Slow Fast 50 2 hours 30 minutes 100 3 hours 45 minutes 250 4 hours 60 minutes 400 4 hours 75 minutes

Three levels were considered for access and egress distances for air, bus, and train, as shown in Table 3.4. The access distance is the distance from the traveler’s home to the airport, bus station, or rail station. The egress distance is the distance from the destination airport, bus station, or rail station to the traveler’s final destination.

Table 3.4 Access and Egress Distance for Air, Bus, and Train Short Medium Long Access Distance 2 miles 10 miles 20 miles Egress Distance 1 mile 5 miles 10 miles

While some other factors may also influence mode choice, such as transfer requirements, the number of factors included in the choice sets was limited to avoid over-burdening survey respondents. As the number of alternatives, attributes, and attribute levels increase, the choice experiment becomes increasingly complex, and the cognitive burden for participants increases. As a result, participants may have difficulty processing all the information and may ignore some attributes, providing less useful results (Hensher 2006, Caussade et al. 2005).

Caussade et al. (2005) found that the two most critical design dimensions were the number of attributes and the number of alternatives. This study and others found a significant negative effect on data quality resulting from an increase in the number of attributes (DeShazo and Fermo 2002, Arentze et al. 2003). Because of the impact of survey complexity on the reliability of results, the number of attributes was limited to those considered most important. Caussade et al. (2005) also found that four alternatives are optimum.

The impact of survey fatigue, resulting from too many choice sets to consider, on data quality is also a potential concern. However, Arentze et al. (2003) did not find any fatigue effect for transportation SP surveys. Caussade et al. (2005) did not find the number of choice sets in the survey to be as important as other design dimensions, but they found that experiments with 9 or 10 choice sets were optimal, minimizing any fatigue effect.

3.1.2.2 Statistical Design

The stated choice experiment consisted of 19 factors with varying levels, as shown in Table 3.5. Three of the factors had two levels, one had four, and the remainder all consisted of three levels. After identifying the factors, selecting the measurement units, and specifying the number and magnitude of factor levels, the next step, following Hensher (1994), was to create the statistical design of the experiment.

Methods from the field of experimental design were used to develop the stated preference questions. Experimental design is a field of statistics concerned with the proper construction of experiments to ensure the preservation of necessary properties. It is often used in marketing research to assist in the investigation of consumer choices. Experimental design methods were used to identify the minimum number of choice sets needed and to construct the individual choice sets.

Table 3.5 Stated Choice Experiment Factors and Levels Number Factor of Levels Levels Distance 4 50 miles, 100 miles, 250 miles, 400 miles Trip type 2 Business, personal Party size 2 Alone, with a group Air price 3 $300/trip, $600/trip, $900/trip Automobile price 3 $2.00/gallon, $3.50/gallon, $5.00/gallon Bus price 3 $0.15/mile, $0.20/mile, $0.25/mile Train price 3 $0.15/mile, $0.20/mile, $0.25/mile Air travel speed 2 Slow, fast Bus travel speed 3 45 mph, 55 mph, 65 mph Train travel speed 3 45 mph, 55 mph, 65 mph Air access distance 3 2 miles, 10 miles, 20 miles Bus access distance 3 2 miles, 10 miles, 20 miles Rail access distance 3 2 miles, 10 miles, 20 miles Air egress distance 3 1 mile, 5 miles, 10 miles Bus egress distance 3 1 mile, 5 miles, 10 miles Rail egress distance 3 1 mile, 5 miles, 10 miles Air frequency 3 Three times per week, once per day, twice per day Bus frequency 3 Three times per week, once per day, twice per day Train frequency 3 Three times per week, once per day, twice per day

Fractional Factorial Main Effects Design

A full-factorial design consists of all possible combinations of the factor levels. Such a design allows the estimation of main effects – simple effects such as the effects of price or travel time – and interactions between variables. A full-factorial design was not feasible for this study, however, as it would result in millions of unique choice sets. To reduce the number of combinations considered, researchers often use fractional-factorial designs. The cost of a fractional-factorial design is that some effects become confounded, or not distinguishable from each other.

Fractional-factorial designs are categorized by their resolution, as explained by Kuhfeld (2010) and shown below:

 Resolution III – All main effects can be estimated free of each other, but some are confounded with two-factor interactions.  Resolution IV – All main effects can be estimated free of each other and free of all two-factor interactions, but some two-factor interactions are confounded with other two-factor interactions.  Resolution V – All main effects and two-factor interactions can be estimated free of each other.

Higher resolutions require larger designs, which are often impractical in marketing or transportation research. As noted by Kuhfeld (2010), Hensher (1994), and Sanko (2001), Resolution III fractional- factorial designs, also referred to as main-effects designs, are commonly used in marketing and transportation research. These designs assume there are no significant interactions between attributes. Main effects and other fractional factorial designs have been shown to be valid, explaining the largest amount of variance in response data, but care should be taken to avoid confounding interaction effects with main effects (Sanko 2001).

Orthogonal and Balanced Design

Fractional-factorial designs that are both orthogonal and balanced are preferred. A design is balanced when each level occurs equally often within each factor, and a design is orthogonal when every pair of levels occurs equally often across all pairs of factors. Orthogonality is important because it ensures that there is no correlation between attributes. While this is a desirable property, it is not a necessary condition (Hensher 1994).

As Kuhfeld (2010) explained, the efficiency of an experimental design can be quantified, and the two most prominent efficiency measures are referred to as the A-efficiency and the D-efficiency. A-efficiency is a function of the arithmetic mean of the eigenvalues, which is also the arithmetic mean of the variances, and D-efficiency is a function of the geometric mean of the eigenvalues. A third efficiency measure, the G-efficiency, is based on the maximum standard error for prediction over the candidate set. The three criteria are usually highly correlated, and they can be scaled to range from 0 to 100. A more efficient design has greater balance and orthogonality. If the design is balanced and orthogonal, then is has optimum efficiency.

According to Kuhfeld (2010), D-efficiency is usually used because it is easier and faster for a computer program to optimize, and the ratio of two D-efficiencies for two competing designs does not vary under different coding schemes, unlike the A-efficiency. Carlsson and Martinsson (2003) also strongly recommend that researchers use a D-optimal design.

Blocking

Even when not using a full-factorial design, the number of choice sets in the design could be too many for the survey respondent. To reduce the choice sets to a reasonable number for the survey participant, the survey questions can be broken into subsets or blocks, which is common in transportation studies (Hensher 1994). Blocking can be successful if the preferences across the samples of respondents receiving different subsets of questions are sufficiently homogenous (Sanko 2001).

SAS Procedures

The statistical design for the SP survey was created with SAS 9.3, using a set of macros described by Kuhfeld (2010). First, the %MktRuns macro was used to choose the number of choice sets. It was found that a balanced and orthogonal design, with optimal D-efficiency, could be achieved with 72 choice sets.

After identifying the optimal number of choice sets, the %MktEx macro was used to efficiently create the choice sets. This macro, which attempts to optimize D-efficiency and provide balanced and orthogonal designs, was created to produce the types of designs that marketing researchers need for choice experiments (Kuhfeld 2010). The resulting design, with 72 choice sets, was found to be sufficient for estimating main effects.

Next, because 72 choice sets are still far too many for an individual survey respondent to complete, the choice sets were assigned to eight blocks, using the %MktBlock macro in SAS, with each block consisting of nine choice sets. The %MktBlock macro attempts to create a blocking factor, which is an additional factor that is uncorrelated with all existing factors and ensures orthogonality within each block.

Final Survey Design

Eight different versions of the stated preference survey were then created with each including a single block of nine choice sets. Each survey respondent, therefore, received nine stated preference questions.

Appendix A shows the resulting survey design. The first table provides the coded survey design, which shows the level for each factor in each of 72 choice sets, as well as the block to which each choice set was assigned. The second table in Appendix A shows the actual factor values for each choice set.

In preparing the survey, prices for bus and train travel were converted from cost per mile to total cost by multiplying the cost per mile by trip distance. Similarly, travel time for automobile, bus, and train were calculated by dividing trip distance by travel speed.

After creating the statistical design, the next step, according to Hensher (1994), was to translate the design into a set of questions. Appendix B provides a final survey design with one of the eight SP blocks.

The final steps outlined by Hensher (1994), selecting an appropriate estimation procedure and using the estimated parameters to obtain choice probabilities, will be discussed in Section 3.3.

3.1.3 Attitudinal Questions

The last section of the survey presented a number of statements about travel, and the respondent was asked to respond on a Likert-type scale the degree to which he or she agreed or disagreed with the statement. These statements, which were used in a previous study by Mattson et al. (2010) and were derived from those used by Outwater et al. (2004), described a number of attitudes regarding the traveler’s sensitivity to the environment, time, flexibility, safety, stress, comfort, reliability, privacy, convenience, and other elements of the travel experience. The full set of attitudinal questions is shown in Appendix B. As shown by Mattson et al. (2010b), an individual’s attitude regarding these aspects of travel can influence choice of mode. Comparing results to previous research will also show if traveler attitudes have changed over time. 3.2 Survey Administration

3.2.1 Study Population, Sample Frame, and Drawing the Sample

The target population for the survey was adults aged 18 or older living in North Dakota. The survey sample was generated from a list of names and addresses obtained from AccuData Integrated Marketing, a marketing firm that sells mailing lists. AccuData maintains frequently updated databases of individual records. These databases are compiled from property data, public records, transactional data, consumer surveys, etc. Most of the records include the age of the individual, as well as a number of different variables. The database from which the sample was drawn contained approximately 220,000 individual records for the state of North Dakota. This represented the sample frame, from which the sample was drawn.

It is desirable for the age and geographic distribution of the sample to resemble those of the study population. The geographic distribution of the sample frame closely resembled the geographic distribution of the study population, as determined by comparing the number of individuals per county. The age distribution of the sample frame, however, was not representative of the study population, as younger adults were underrepresented in the AccuData database. Therefore, a stratified sampling technique was used, where the sample frame was divided into subsets, based on age, and a random sample was drawn from each subset. When purchasing the records, a specific number of records was requested

18 from each age group, so the age distribution of the sample resembled that of the study population. Because these records were randomly chosen, the geographic distribution of the sample should be similar to that of the study population.

The sample frame was divided into two-year age bands. Table 3.6 shows the percentage of the North Dakota population belonging to each of these age bands (according to 2014 population estimates from the Census), the number of individual records purchased from each age band, and the percentage of the sample belonging to each age band. Because of the very low number of individuals aged 18-19 in the AccuLeads database, it was not possible to obtain proportional representation of this age group in the sample. All other age groups, however, were represented appropriately.

Table 3.6 Age Distribution of Survey Sample and North Dakota Population

Number in Percentage of Percentage of North Age Sample Sample Dakota Population 18-19 114 2.3% 4.0% 20-21 244 4.9% 4.8% 22-23 269 5.4% 5.3% 24-25 232 4.6% 4.5% 26-27 194 3.9% 3.8% 28-29 197 3.9% 3.9% 30-31 190 3.8% 3.7% 32-33 184 3.7% 3.6% 34-35 171 3.4% 3.4% 36-37 156 3.1% 3.1% 38-39 144 2.9% 2.8% 40-41 136 2.7% 2.7% 42-43 142 2.8% 2.8% 44-45 145 2.9% 2.8% 46-47 140 2.8% 2.7% 48-49 151 3.0% 3.0% 50-51 168 3.4% 3.3% 52-53 179 3.6% 3.5% 54-55 184 3.7% 3.6% 56-57 180 3.6% 3.5% 58-59 175 3.5% 3.4% 60-61 163 3.3% 3.2% 62-63 149 3.0% 2.9% 64-65 132 2.6% 2.6% 66-67 125 2.5% 2.4% 68-69 94 1.9% 1.8% 70-71 91 1.8% 1.8% 72-73 77 1.5% 1.5% 74-75 68 1.4% 1.3% 76-77 65 1.3% 1.3% 78-79 61 1.2% 1.2% 80-81 56 1.1% 1.1% 82-83 52 1.0% 1.0% 84+ 172 3.4% 3.6%

3.2.2 Sample Size

Suzuki et al. (2002) studied design issues in a stated preference survey for a light rail project in Japan, including the effects of sample size on mode split models, and they found that it is possible to estimate a well-behaved model from 100-300 samples if the survey instrument is well designed and administered. Other transportation studies have estimated models using SP survey data with a similar number of responses, such as Kumar et al. (2004), Andrade et al. (2006), and Srinivasen et al. (2006). While 100-300 usable SP responses appears sufficient and consistent with other published transportation research, a review of stated preference methods published by Accent and RAND Europe (2010) recommended a minimum of 400 survey responses for market research. Therefore, the goal for the survey was 400-500 responses. This number of responses would provide a 5% sampling error with a 95% confidence level.

The study employed a combination of mail and online surveys and, based on expected response rates for these types of surveys, a sample of 5,000 individuals was determined adequate for obtaining 400-500 responses (assuming that about 5% of individual records would be outdated, that the response rate for mail surveys would be about 12%-20%, and that the response rate for online surveys would be even lower).

3.3.3 Survey Distribution To reduce costs, half of the surveys were administered online. Individuals in the sample were randomly assigned to either the mail survey or the online survey, while maintaining a similar age distribution between the two groups. When assigning individuals to the mail or online survey, potential survey participants were categorized into four age groups: 18-29, 30-49, 50-69, and 70 or older. In each of these four age groups, half of the individuals were randomly chosen to receive the paper survey, and the other half received the online survey. This procedure allowed for a similar age distribution between recipients of the two types of surveys.

Since the SP questions were blocked into eight groups, there were eight different versions of the survey instrument. Those selected to receive the mail survey were randomly assigned to receive one of the eight versions, with an equal number of recipients of each (four versions of the survey each had 313 recipients and the other four versions each had 312 recipients, totaling 2,500 recipients). The online survey was developed so that participants would randomly be given one of the eight SP blocks.

Recipients of the mail survey received a mailing consisting of a cover letter, the survey, and a postage- paid return envelope. Recipients of the online survey received a postcard in the mail with information about the survey and a web address where they can access and participate in the survey. Recipients of the mail survey were also given the web address so they could take the survey online if they preferred.

One to two weeks after the initial mailing, a postcard reminder was sent to all potential survey participants. To boost the response rate further, a cash incentive was used. All survey participants who provided their name and contact information were entered into a drawing to win a $300 cash prize. While research tends to show that these types of lottery incentives are not effective for increasing response rates for mail surveys, some studies have found that they can be quite effective for web-based surveys (Heerwegh 2006, Laguilles et al. 2011). 3.3 Discrete Choice Modeling

The survey data were analyzed with a discrete choice model. Factors affecting mode choice can be organized into three categories: the characteristics of the different transportation modes, the characteristics of the individual making the trip, and the characteristics of the trip itself. A type of discrete choice model was developed based on this framework, using data collected from the survey. The model

20 estimated the probability of a survey respondent choosing a given mode as a function of the mode characteristics (e.g., cost, travel time, service frequency, etc.), trip characteristics (e.g., trip purpose, party size), and individual characteristics (e.g., age, income, gender, etc.). Results from the model can be used to predict the probability that an individual with a given set of characteristics will choose a certain mode with given mode attributes for a certain type of trip. 3.3.1 Multinomial Logit Model

Discrete choice modeling is popular in transportation and marketing research for understanding an individual’s stated choice among alternatives (Kuhfeld 2010). The multinomial logit model, which is a type of random utility choice model, has been traditionally used to model the choice among alternative modes in intercity travel demand modeling (Koppelman and Sethi 2005).

The basic assumption in such a model is that decision makers are utility maximizers. Therefore, given a set of alternatives, the decision maker will select the one that maximizes his or her utility, Ujk. The utility of an alternative k for decision maker j is assumed to consist of a deterministic part that can be estimated,

Vjk, and a random portion called the error term, jk, as follows:

Ujk = Vjk + jk (1)

The deterministic part of the utility function includes characteristics of the decision maker, trip characteristics, and characteristics of the alternatives. Different assumptions about the distribution of the error component result in different choice models. The simplest and most widely used among the multinomial discrete choice models is the conditional logit, which is often what is being referred to as the multinomial logit model (Koppelman and Sethi 2005).

The deterministic portion of the utility from Equation 1 can be written as follows:

Vjk = Xj + Yj + Zjk (2) where Xj are the characteristics of individual j; Yj are the trip characteristics for individual j; Zjk are the attributes of alternative k for individual j; and , γ, and  are the parameter vectors associated with the vectors Xj, Yj, and Zjk, respectively. The probability that individual j would choose mode k among m alternatives is as follows:

exp⁡(훽푘푋푗+푘푌푗⁡+⁡휃푍푗푘) (3) 푃푗푘 = 푚 ⁡. ∑푙=1 exp⁡(푙푋푗+푙푌푗⁡+⁡휃푍푗푙)

The data collected from the stated preference survey were used to estimate , γ, and . Discrete choice models usually use Maximum Likelihood Estimation (MLE) to estimate these values. Each choice set completed by a survey participant is represented in the data by a block of four rows (one for each mode), columns for each observable characteristic (Xj, Yj, and Zjk), and a column indicating if the mode was chosen. MLE is used to estimate values of , γ, and  that make the observed choices have the highest probability. As described by Ortuzar and Willumsen (2001), maximum likelihood estimates are the set of parameters which will generate the observed sample most often.

A problem with the multinomial logit model is that it suffers from the Independence from Irrelevant Alternatives (IIA) property, which makes the model inappropriate for situations where some pairs of alternatives are more substitutable than others (Koppelman and Sethi 2005). The restriction imposed by the IIA property implies equal competition between all pairs of alternatives. Therefore, the introduction of

21 a new mode or improvements to an existing mode will reduce the probabilities for existing modes in proportion to their probabilities before the change (Koppelman and Bhat 2006). In reality, some modes are more similar to each other than they are to other alternatives. For example, for intercity travel, bus and rail modes have shared attributes which are not included in the measured portion of the utility function, as they are both forms of public transportation with lack of privacy, lack of control of the environment, etc. As a result, they may be more substitutable than others, so improvements to bus service would disproportionately affect the rail mode share.

3.3.2 Nested Logit Model Hensher and Greene (2002) argued that the nested logit model is the preferred specification of a discrete choice model when moving beyond the multinomial logit model, and Forinash and Koppelman (1993) found nested logit structures for intercity mode choice modeling to be superior to the multinomial logit model. As Wen and Koppelman (2001) noted, the nested logit model is the most widely known relaxation of the multinomial logit model. An advantage of the nested model is that similar modes can be grouped as a subset.

Figure 3.1 is an example of a two-level nest structure that could be used for modeling intercity mode choice. The figure depicts an upper-level choice among automobile, air, and public transit and a lower- level choice between bus and rail, given that public transit is chosen.

Public Transit

Automobile Air Bus Rail Figure 3.1 Two-Level Nest Structure

Nested logit probabilities can be expressed as the product of two simple logits (UC Berkeley 2000).

P = Prob (nest containing i) x Prob (i, given nest containing i) (4)

So,

P(Bus) = Prob (PT) x Prob (Bus|PT) (5) Following Koppelman and Bhat (2006), the choice probabilities for the lower-level nested alternatives (bus or rail), conditional on choice of these alternatives, are given by: 푉 푒푥푝 ( 퐵푢푠) (6) 휃 Pr(퐵푢푠|푃푇) = ⁡ 푃푇 푉 푉 푒푥푝 ( 퐵푢푠) + 푒푥푝 ( 푅푎푖푙) 휃푃푇 휃푃푇

푉 (7) 푒푥푝 ( 푅푎푖푙) 휃 Pr(푅푎푖푙|푃푇) = ⁡ 푃푇 푉 푉 푒푥푝 ( 푅푎푖푙) + 푒푥푝 ( 퐵푢푠) 휃푃푇 휃푃푇

휃푃푇 is the logsum parameter, bounded by zero and one. The logsum parameter measures how similar items are within a nest. The marginal choice probabilities for the automobile, air, and public transit alternatives are:

exp⁡(푉퐴푢푡표 ) (8) Pr(퐴푢푡표) = ⁡ exp(푉퐴푢푡표) + exp(푉퐴푖푟) + exp⁡(푉푃푇 + 휃푃푇Γ푃푇)

exp⁡(푉퐴푖푟) (9) Pr(퐴푖푟) = ⁡ exp(푉퐴푢푡표) + exp(푉퐴푖푟) + exp⁡(푉푃푇 + 휃푃푇Γ푃푇)

exp⁡(푉푃푇 + 휃푃푇Γ푃푇) (10) Pr(푃푇) = ⁡ exp(푉퐴푢푡표) + exp(푉퐴푖푟) + exp⁡(푉푃푇 + 휃푃푇Γ푃푇)

Γ푃푇 represents the expected value of the maximum of the bus and rail utility and is computed as follows:

푉퐵푢푠 푉푅푎푖푙 (11) Γ푃푇 = 푙표푔 |푒푥푝 ( ) + 푒푥푝 ( )| 휃푃푇 휃푃푇 3.3.3 Mixed Logit Model

The mixed logit model is another alternative to the multinomial logit model. Like the nested logit model, the mixed logit model does not suffer from the IIA property (Hensher and Greene 2003). With improvements in software packages and the development of simulation methods, the mixed logit model became more popular in the 1990s and 2000s, when it became considered one of the more promising state-of-the-art discrete choice models available (Hensher and Greene 2003).

The mixed logit model models heterogeneity of multiple choices, with the probability of choosing alternative j written as

′ exp⁡(푥푖푗훽) 푃 (푗) = ⁡ (12) 푖 퐽 ′ ∑푘=1 exp⁡(푥푖푘훽) where  is allowed to vary randomly (SAS 2010). This model uses a Monte Carlo simulation method to estimate the probabilities of each choice. The mixed logit model takes into account that the error components of different alternatives could be correlated by partitioning the stochastic component into two parts

′ 푈푖푗 = 푥푖푗훽 + [휂푖푗 + 휀푖푗] (13) where 휂푖푞 can be correlated among alternatives and heteroscedastic for each individual while 휀푖푗 is independently and identically distributed.

3.3.4 Determining Appropriate Model

The multinomial logit model can be used if the IIA property is not problematic. The Hausman specification test can be conducted to analyze the IIA assumption (SAS 2010). If this test rejects the null hypothesis that the IIA property holds, then a different model that does not assume the IIA property would be more appropriate. The appropriateness of the nested logit model can be determined based on the estimate of the logsum parameters (), which measure how similar items are within a nest. If the estimated logsum parameter is statistically insignificant, then the nest is not necessary.

To estimate the mixed logit model, the random parameters must be selected. One method for selecting random parameters, as suggested by Hensher and Greene (2003) is to first assume all parameters are random and then examine their estimated standard deviations using a t-test. Mariel et al. (2011) noted that this is a common procedure for most applications of discrete choice experiments in the literature. 3.4 Statewide Travel Demand Model

The estimated mode choice model can be incorporated into a statewide travel demand model to estimate demand for each mode. Such a model estimates an origin-destination (O-D) matrix of trips across the state. By adding the intercity bus network to the model and including data for travel time, access and egress distances, fares, and service frequency, as well as data for other modes, the number of trips taken by bus for each O-D pair can be estimated.

Mode shares for each O-D pair were estimated using travel time for each mode; access and egress distances for bus, rail, and air; current costs for each mode; and service frequency for bus, rail, and air. To account for demographic impacts on mode choice, demographic data for each TAZ were obtained from the American Community Survey.

This model could then be used to estimate the impact of hypothetical changes to the intercity bus network, changes in the service characteristics of intercity bus, changes in the attributes of competing modes, or demographic changes in the state.

4. SURVEY RESPONSE AND SUMMARY STATISTICS

4.1 Response Rate

Among the 2,500 paper surveys sent by mail, 226 were returned undeliverable due to incorrect addresses, which left 2,274 surveys that were presumably delivered. A total of 393 responses were received, yielding a response rate of 17.3% (Table 4.1). Recipients of the paper survey had the option of completing the paper survey and returning it by mail or completing it online. Among the 393 responses, 52 were completed online, and 341 were paper responses delivered by mail.

Table 4.1 Survey Response Rate Paper + Online Online Only Total Sample 2,274 2,489 4,763 Responses* 393 148 541 Response Rate 17.3% 5.9% 11.4% *For the paper+online survey, 341 paper surveys were completed, and 52 were received online

A second group of 2,500 potential participants received a postcard informing them of the survey with instructions on how to complete the survey online. Recipients were directed to a website where they could participate. Among the 2,500 postcards sent, 11 were returned undeliverable. A total of 148 online responses were received, yielding a 5.9% response rate.

Combined, 541 responses were received from the 4,763 paper surveys or postcards that were delivered. The overall response rate was 11.4%. The response rates were within the range of expectations. The number of responses received per block ranged from 56 to 76. 4.2 Summary Statistics

4.2.1 Demographic Characteristics of Respondents

Table 4.2 shows the demographic characteristics of survey respondents. Differences are shown between the paper survey and online survey respondents, and the demographic characteristics of the target population, which is North Dakota adults aged 18 or older, is also presented. Overall, the survey respondents were fairly representative of the North Dakota population. Men were slightly overrepresented, as 55% of respondents were male, compared to 51% of the target population. Young adults under age 25 were underrepresented, while the 25-69 age groups were slightly overrepresented, and adults 80 or older were underrepresented. The gender and age distributions of the survey respondents appear reasonable, though, as men tend to drive more miles per year than women, and those aged 25-69 also drive more miles (Mattson 2012).

Two measures of geographic distribution are illustrated in Table 4.2. First, the state was divided into nine regions based on the first three digits of the zip code, as shown in Figure 4.1. The percentage of respondents from each of these regions closely matched the percentage of the state’s population in each region, showing that all regions of the state were adequately represented.

Table 4.2 Demographic Characteristics of Survey Respondents and Target Population Target Paper Online Total Population* (No.) (%) (No.) (%) (No.) (%) (%) Gender Male 195 58 102 51 297 55 51 Female 144 42 98 49 242 45 49 No response 2 0 2 Age <25 29 9 23 12 52 10 16 25-29 31 9 39 20 70 13 10 30-39 56 17 50 25 106 20 16 40-49 34 10 29 15 63 12 14 50-59 70 21 31 16 101 19 17 60-69 67 20 23 12 90 17 13 70-79 37 11 4 2 41 8 7 80+ 14 4 1 1 15 3 6 No response 3 0 3 3-Digit Zip Code 580 56 16 31 16 87 16 12 581 52 15 33 17 85 16 16 582 45 13 35 18 80 15 14 583 22 6 8 4 30 6 7 584 24 7 13 7 37 7 7 585 61 18 48 24 109 20 20 586 18 5 7 4 25 5 6 587 35 10 18 9 53 10 13 588 7 2 3 2 10 2 4 No response 21 4 25 Urban/Rural Metropolitan City 140 44 112 57 252 49 44 Micropolitan City 49 15 27 14 76 15 18 Towns 2,500-10,000 22 7 9 5 31 6 5 Rural Areas <2,500 109 34 49 25 158 31 32 No response 21 3 24 Household Income < $25,000 44 13 21 11 65 12 22 $25,000-$49,999 69 21 42 22 111 21 25 $50,000-$74,999 84 25 34 17 118 22 19 $75,000-$99,999 56 17 33 17 89 17 14 $100,000+ 80 24 65 33 145 27 20 No response 8 5 13 *Sources: American Community Survey 2014 1-Year Estimate, 2009-2013 5-Year Estimates

Figure 4.1 North Dakota Regions based on Three-Digit Zip Codes

Second, based on the zip codes provided by respondents, each individual was classified as living in a metropolitan city, a micropolitan city, a town with population of 2,500 to 10,000, or a rural area, as shown in Table 4.3. There are three metropolitan areas in North Dakota and five micropolitan areas. A respondent was classified as living in a metropolitan city, micropolitan city, or town with a population of 2,500 to 10,000 if his or her zip code belonged to one of the cities shown in Table 4.3.1 All other zip codes were classified as rural.

Based on these classifications, 49% of respondents were from a metropolitan city, 15% were from a micropolitan city, 6% were from a town with a population of 2,500 to 10,000, and 31% were from rural areas. This again shows a good geographic distribution of survey respondents, similar to the statewide population distribution, though residents of metropolitan cities were somewhat overrepresented.

1 Metropolitan areas are defined as having an urban core area population of 50,000 or more, and micropolitan areas are defined as having an urban core population of 10,000 to 50,000. While Wahpeton has a population of less than 10,000, it combines with Breckinridge, MN, to have an urban core population greater than 10,000. For the classification purposes of this study, only the urban core cities are classified as metropolitan or micropolitan.

Table 4.3 Geographic Classification of North Dakota Areas Metropolitan City Micropolitan City Town 2,500 to 10,000 Rural Population Fargo Minot Devils Lake All others West Fargo Williston Valley City Horace Dickinson Grafton Bismarck Jamestown Watford City Mandan Wahpeton Beulah Lincoln Rugby Grand Forks Casselton

Lastly, Table 4.2 shows the income distribution of survey respondents. While middle-income residents tended to be fairly adequately represented, low-income residents were underrepresented in the survey and high-income residents were overrepresented (12% of respondents had a household income of less than $25,000, compared to the statewide total of 22%). Again, this may be reasonable as higher-income individuals tend to make more trips overall (Mattson 2012).

There are some demographic differences between the paper survey respondents and the online survey respondents. Notably, the online respondents skewed younger, as might be expected. Fifty-six percent of online respondents were younger than 40, and just 14% were 60 or older. Conversely, 34% of paper survey respondents were under the age of 40 and 35% were 60 or older. In addition to being younger, online respondents were more likely to be from a metropolitan area and to have a higher income.

4.2.2 Transportation Characteristics of Respondents

Nearly all survey respondents reported being able to operate an automobile, and just a small percentage reported having a medical condition or disability that makes it difficult to travel (Figure 4.2). More precisely, 34 out of 540 respondents, or 6%, reported having a medical condition or disability that makes it difficult to travel, and just eight respondents, or 1.5%, reported being unable to drive an automobile. Only 1% of respondents were from households without any automobile, while close to four out of five had two or more automobiles in their household. Most respondents (91%) had made a long-distance trip of 100 miles or more one-way within the previous year, though 9%had not taken any such trips. Most respondents were not recent users of intercity bus or rail services, but many had traveled by airplane. Within the previous five years, 15% had traveled by intercity rail, 13% had traveled by intercity bus, and 79% had traveled by airplane.

Do you have a medical condition or disability that makes it difficult to travel? Yes 6% No 94%

Are you able to operate an automobile (legally, physically, mentally)?

Yes 98.5% No 1.5%

How many automobiles are kept at home for use by members of this household?

None 1% One 20% Two 43% Three or more 36%

During the past year, did you make any long-distance trips of 100 miles or more one-way?

Yes 91% No 9%

Have you traveled by intercity rail anywhere within the last five years?

Yes 15% No 85%

Have you traveled by intercity bus anywhere within the last five years?

Yes 13% No 87%

Have you traveled by airplane anywhere within the last five years?

Yes 79% No 21%

Figure 4.2 Transportation Characteristics of Survey Respondents

4.2.3 Opinions on Ease of Travel and Need for Improvements

Most respondents reported that it was easy for them to travel to other cities in North Dakota. Of 539 responses, 73% answered that it was either easy or very easy (Figure 4.3). Ten percent reported that it was either somewhat difficult, difficult, or very difficult.

Very easy 40%

Easy 33%

Somewhat easy 10%

Neither difficult nor easy 8%

Somewhat difficult 6%

Difficult 2%

Very difficult 2%

Figure 4.3 Survey Responses on How Easy it is to Travel to Other Cities in North Dakota

Survey respondents were then asked, regarding travel between towns and cities in North Dakota, how much there is a need for highway improvements, bus service improvements, passenger rail service improvements, and air service improvements (Figure 4.4). Respondents were most likely to indicate a need for highway improvements, which was not surprising given that the automobile was the predominant mode of travel within the state. Twenty percent indicated a high need for highway improvements, and 55% answered there is a moderate need.

Many answered (about half of respondents) that they do not know if there is a need for bus or rail service improvements, which again was not surprising given the lower usage of these modes. Despite the low use of bus and rail services by these respondents, many indicated a need for service improvements. Twenty- nine percent said there was a high or moderate need for bus service improvements, and 34% answered the same for rail service improvements. Very few answered that there was no need for improvements of bus or rail. Regarding air service, 46% answered there was a high or moderate need for improvements.

Highway Improvements Bus Service Improvements

High need 20% High need 11%

Moderate need 55% Moderate need 18%

Low need 22% Low need 15%

No need at all 2% No need at all 4%

Do not know 2% Do not know 52%

Passenger Rail Improvements Air Service Improvements High need 16% High need 13%

Moderate need 17% Moderate need 33%

Low need 16% Low need 25%

No need at all 5% No need at all 8%

Do not know 46% Do not know 22%

Figure 4.4 Survey Responses on Need for Intercity Transportation Improvements in North Dakota

4.2.4 Opinions on Travel

The average responses from the opinion questions on travel preferences and attitudes are shown in Table 4.4. Responses were given on a 1-10 scale, with a higher number indicating greater agreement. The highest rated statement, with an average response of 8.4, was, “If my travel options are delayed, I want to know the cause and length of the delay.” In fact, 42% of respondents answered with a 10 on this statement. The next highest rated statements were regarding timeliness, comfort, predictability, and cleanliness.

Environmental issues were not a primary concern for survey respondents. The lowest-rated statement, with an average score of 2.7, was, “People who travel alone should pay more to help improve the environment.” In fact, 45% gave a response of 1 to this statement. Statements regarding the willingness to pay more or to switch to a different form of transportation if it would help improve the environment also received among the lowest scores.

The average response to the statements that people who fly, ride intercity rail, or use intercity bus services are “like me” were 5.6, 4.5, and 4.1, respectively. The lower responses for intercity rail and, especially, bus suggest that people perceive users of these services to be different from them.

Other statements receiving low ratings regarded concerns about getting into an accident (respondents tended to be not too worried) and long delays (although they valued comfortable seats, the statement “I don’t mind long delays as long as I’m comfortable” received a low score).

Table 4.4 Response to Opinion Questions Average Score Statement 8.4 If my travel options are delayed, I want to know the cause and length of the delay. When traveling, I like to keep as close as possible to my departure and arrival 8.1 schedules. 8.0 It is important to have comfortable seats when I travel. 8.0 I prefer a travel option that has a predictable travel time. 7.8 A clean vehicle is important to me. 7.0 I would like to make productive use of my time when traveling. 6.8 I would change my form of travel if it would save me some time. 6.8 Having a stress-free trip is more important than reaching my destination quickly. 6.2 I would rather do something else with the time that I spend traveling. 5.9 Having privacy is important to me when I travel. 5.7 It’s important to be able to change my travel plans at a moment’s notice. 5.6 I need to make trips according to a fixed schedule. 5.6 I avoid traveling at certain times because it is too stressful. 5.6 The people who fly are like me. 5.5 I prefer to make trips alone, because I like the time to myself. 5.5 When traveling, I like to talk and visit with other people. 5.4 I’m willing to pay more for a ticket if it allows me to re-book my trip later for free. 5.0 I use the most convenient form of transportation regardless of cost. 5.0 I always take the fastest route to my destination even if I have a cheaper alternative. 4.7 I don’t mind traveling with strangers. 4.5 The people who use intercity rail service are like me. 4.5 I would switch to a different form of transportation if it would help the environment. 4.4 I worry about getting in an accident when I travel. 4.4 I don’t mind long delays as long as I’m comfortable. 4.1 The people who ride intercity bus are like me. 4.1 I would be willing to pay more when I travel if it would help the environment. 2.7 People who travel alone should pay more to help improve the environment.

Responses to these attitudinal questions were very similar to those from a previous survey conducted in 2009 (Mattson et al. 2010b). This survey asked the same attitudinal questions of residents from North Dakota and western Minnesota. The differences in responses between the two surveys are largely negligible, as shown in Table 4.5. A t test was conducted to measure the statistical significance of the differences. Most differences were found to be statistically insignificant. A few were found to be statistically significant, but these differences are small in magnitude. There does not appear to be any trend or change in traveler attitudes in North Dakota during this period.

Table 4.5 Differences in Traveler Attitudes from 2009 to 2015 Average Score Difference Statement 2015 2009 5.6 5.0 0.6** The people who fly are like me. 6.8 6.3 0.5** I would change my form of travel if it would save me some time. 7.0 6.5 0.5** I would like to make productive use of my time when traveling. 4.5 4.1 0.4* The people who use intercity rail service are like me. 4.4 4.0 0.4* I worry about getting in an accident when I travel. 6.2 5.8 0.4 I would rather do something else with the time that I spend traveling. 8.0 7.6 0.4* I prefer a travel option that has a predictable travel time. 4.1 3.8 0.3 The people who ride intercity bus are like me. 5.6 5.4 0.2 I need to make trips according to a fixed schedule. If my travel options are delayed, I want to know the cause and length of 8.4 8.2 0.2 the delay. When traveling, I like to keep as close as possible to my departure and 8.1 7.9 0.2 arrival schedules. I’m willing to pay more for a ticket if it allows me to re-book my trip 5.4 5.2 0.2 later for free. 5.5 5.4 0.1 I prefer to make trips alone, because I like the time to myself. I would switch to a different form of transportation if it would help the 4.5 4.4 0.1 environment. Having a stress-free trip is more important than reaching my destination 6.8 6.7 0.1 quickly. 4.4 4.3 0.1 I don’t mind long delays as long as I’m comfortable. 7.8 7.7 0.1 A clean vehicle is important to me. I always take the fastest route to my destination even if I have a cheaper 5.0 5.0 0.0 alternative. 5.7 5.7 0.0 It’s important to be able to change my travel plans at a moment’s notice. I would be willing to pay more when I travel if it would help the 4.1 4.1 0.0 environment. 8.0 8.1 -0.1 It is important to have comfortable seats when I travel. 5.9 6.0 -0.1 Having privacy is important to me when I travel. 5.5 5.6 -0.1 When traveling, I like to talk and visit with other people. 5.0 5.2 -0.2 I use the most convenient form of transportation regardless of cost. 4.7 4.9 -0.2 I don’t mind traveling with strangers. 5.6 6.0 -0.4 I avoid traveling at certain times because it is too stressful. People who travel alone should pay more to help improve the 2.7 3.1 -0.4 environment. * denotes difference is statistically significant at p<0.05 ** denotes difference is statistically significant at p<0.01

4.2.5 Stated Preference Response

With each survey respondent given nine different SP questions to answer, there were a total of 4,724 responses received. The automobile was the mode of choice in 73% of these responses, while air, rail, and bus accounted for 13%, 10%, and 4% of responses, respectively (Figure 4.5). While it was expected that the automobile would be the dominant mode of choice, the results also revealed a clear preference for rail travel over bus travel. More interesting than the overall mode shares is how mode shares change with changes in demographic, trip, and mode characteristics.

Automobile 73%

Air 13%

Rail 10%

Bus 4%

Figure 4.5 Mode Choice Responses from Stated Preference Survey

Table 4.6 shows changes in mode choices with changes in demographic characteristics. Men were more likely than women to choose automobile, and a greater percentage of women chose rail travel. Older adults (age 70 or older) and younger adults (age 18-24) were more likely to choose bus travel. Older adults tended to be less likely to choose air travel, choosing automobile travel instead. Those from lower income households tended to be less likely than others to choose automobile or air travel and more likely to choose bus or rail travel. For example, bus travel was chosen 6% of the time by those with income below $50,000 but just 3% of the time by those with income of $75,000 or more. Similar differences were found for rail travel.

Table 4.6 Mode Choice Data from the Stated Preference Survey by Individual Characteristics Auto Air Bus Rail Characteristic (%) (%) (%) (%) Gender Male 75 12 4 8 Female 70 14 4 12 Age < 25 71 13 6 10 25-49 70 16 4 10 50-69 77 11 4 9 70+ 76 7 6 11 Household Income <$25,000 69 11 6 14 $25,000-$49,999 67 14 6 12 $50,000-$74,999 78 10 4 8 $75,000-$99,999 74 14 3 9 $100,000+ 74 14 3 9

The age composition of respondents who chose each mode is presented in Table 4.7. Among those who chose the bus, 13.8% were aged 18-24 and 14.3% were aged 70 or older. Comparatively, among all SP survey responses, 9.7% were from those aged 18-24 and 10.4% were from those 70 or older. This result shows bus travel having a larger share of young users and old users, compared to other modes, though the difference may not be as great as shown in some surveys of intercity bus users. The table also shows that just 5.9% of those who chose bus were 70 or older.

Table 4.7 Age Composition by Mode Chosen in Stated Preference Survey Age <25 25-69 70+ (%) (%) (%) Mode chosen Automobile 9.5 79.6 10.9 Air 9.8 84.3 5.9 Bus 13.8 71.9 14.3 Rail 10.1 78.5 11.4 All respondents 9.7 79.9 10.4

Mode choice percentages changed significantly for automobile, air, and rail travel based on the length of the trip, as shown in Table 4.8. Auto shares decreased with increases in travel distance, while air and rail shares increased. While air travel was rarely chosen for 50- or 100-mile trips, it was chosen 12% of the time for 250-mile trips and 37% of the time for 400-mile trips. Meanwhile, rail shares increased from 6% for 50-mile trips to 14% for 400-mile trips. Bus shares, on the other hand, were consistently 4-5%, regardless of trip distance.

Table 4.8 Mode Choice Data from the Stated Preference Survey by Trip Characteristics Auto Air Bus Rail Characteristic (%) (%) (%) (%) Trip distance 50 miles 89 1 4 6 100 miles 85 2 4 8 250 miles 71 12 4 12 400 miles 45 37 5 14 Trip purpose Personal 77 11 3 9 Business 69 15 5 11 Party Size Alone 71 13 5 11 Group 75 13 3 8

Respondents were more likely to choose automobile travel for personal trips than for business trips, while air, bus, and rail travel were chosen more often for business trips. These results are consistent with previous findings by Mattson et al. (2010b). Survey findings also showed that those traveling alone were less likely to choose the automobile and more likely to choose bus and rail travel. This finding is reasonable as there is more of a cost disadvantage for groups to travel by bus or rail (as well as air) because each traveler would need to purchase a ticket, whereas they could share the cost of automobile travel.

The impacts of mode characteristics on mode choices are shown in Table 4.9. Of note are the changes in mode choices as the price of gasoline changes. As the price of gasoline increases from $2.00 to $5.00, the auto share decreases from 79% to 65%, while the bus share increases from 3% to 6% and the rail share increases from 6% to 15%. The most significant impacts occur as the price of gasoline increases from $3.50 to $5.00. The impacts on mode shares are much smaller when the gasoline price varies between $2.00 and $3.50.

Table 4.9 Mode Choice Data from the Stated Preference Survey by Mode Characteristics Auto Air Bus Rail Characteristic (%) (%) (%) (%) Price of gasoline $2.00/gallon 79 12 3 6 $3.50/gallon 75 14 3 8 $5.00/gallon 65 14 6 15 Trip cost Low 79 16 5 13 Medium 75 13 5 9 High 65 10 3 8 Speed Slow na 10 4 7 Medium na na 4 11 Fast na 16 5 11 Access distance 2 miles na 12 5 11 10 miles na 13 4 10 20 miles na 13 4 8 Egress distance 1 mile na 13 5 11 5 miles na 12 5 9 10 miles na 14 3 10 Frequency Three times per week na 14 4 10 Once per day na 12 5 10 Twice per day na 13 5 9 na = not applicable

The impact of trip cost demonstrates the changes in mode shares for each mode as the cost of that mode changes from low to medium to high. For example, bus travel was chosen 5% of the time when the bus price was low and 3% of the time when the bus price was high. Three price levels were surveyed for each mode, as shown previously in Table 3.2. The results are as expected, with each mode being chosen less often as its cost increases.

Similarly, Table 4.9 shows changes in mode shares for each mode with changes in its own speed, access and egress distance, and service frequency. As expected, each mode is chosen more often when it travels at a higher speed. Respondents tended to be more likely to choose bus and rail travel as access and egress distances for those modes decreased, though access and egress distances do not appear to have any impact on choice of air travel. Lastly, service frequency did not appear to influence choice of air or rail travel, while there was a slight increase in bus shares as bus frequency increased.

These statistics are based on the raw survey data and are presented to illustrate correlations found in the survey responses. To determine the statistical significance of these relationships and disentangle the various factors that could be impacting mode choice, these data are analyzed using a discrete choice mode

37 choice model, as described in Section 3.3. The model specification is detailed in Section 5 and results are presented in Section 6.

Further, the results from the survey are all based on the hypothetical conditions presented in the stated choice experiments and may not be representative of real-world conditions present to North Dakota travelers. For example, many North Dakotans do not live within 20 miles of a rail or bus station, so the actual, real-world mode share for bus and rail travel are expected to be smaller. Real-world conditions will be considered when the mode choice model is incorporated into a statewide travel demand model, which will estimate actual access and egress distances and travel times for each mode. This incorporation into the statewide model is presented in Section 7.

5. MODEL SPECIFICATION

Two models were estimated. The first was a mixed logit mode choice model designed to be incorporated into a statewide travel demand model. The second was a binary logit model used to examine the impacts of attitudes on choice of mode. 5.1 Mixed Logit Mode Choice Model

As discussed in Section 3, the deterministic portion of the utility is written as follows:

Vjk = Xj + Yj + Zjk (14) where Xj are the characteristics of individual j; Yj are the trip characteristics for individual j; Zjk are the attributes of alternative k for individual j; and , γ , and  are the parameter vectors associated with the vectors Xj, Yj, and Zjk, respectively. More specifically, utility for each mode is specified as follows:

푉 = ⁡ 훽 ∙ 푥 + 훽 ∙ 푥 + 훽 ∙ 푥 + 훽 ∙ 푥 + 훽 ∙ 푥 푎푢푡표 1푎푢푡표 1푎푢푡표 2푎푢푡표 2푎푢푡표 3푎푢푡표 3푎푢푡표 4푎푢푡표 4푎푢푡표 5푎푢푡표 5푎푢푡표 (15) + 훾1푎푢푡표 ∙ 푦1푎푢푡표 + 훾2푎푢푡표 ∙ 푦2푎푢푡표 + 휃1푎푢푡표 ∙ 푧1푎푢푡표 + 휃2 ∙ 푧2푎푢푡표 + 휃3푚 ∙ 푧3푎푢푡표

푉 = ⁡ 훽 ∙ 푥 + 훽 ∙ 푥 + 훽 ∙ 푥 + 훽 ∙ 푥 + 훽 ∙ 푥 + 훾 ∙ 푦 + 훾 푎푖푟 1푎푖푟 1푎푖푟 2푎푖푟 2푎푖푟 3푎푖푟 3푎푖푟 4푎푖푟 4푎푖푟 5푎푖푟 5푎푖푟 1푎푖푟 1푎푖푟 2푎푖푟 (16) ∙ 푦2푎푖푟 + 휃1푎푖푟 ∙ 푧1푎푖푟 + 휃2 ∙ 푧2푎푖푟 + 휃3 ∙ 푧3푎푖푟 + 휃4 ∙ 푧4푎푖푟 + 휃5 ∙ 푧5푎푖푟 + 휃6 ∙ 푧6푎푖푟 + 휃7 ∙ 푧7푎푖푟

푉 = ⁡ 훽 ∙ 푥 + 훽 ∙ 푥 + 훽 ∙ 푥 + 훽 ∙ 푥 + 훽 ∙ 푥 + 훾 ∙ 푦 푏푢푠 1푏푢푠 1푏푢푠 2푏푢푠 2푏푢푠 3푏푢푠 3푏푢푠 4푏푢푠 4푏푢푠 5푏푢푠 5푏푢푠 1푏푢푠 1푏푢푠 (17) + 훾2푏푢푠 ∙ 푦2푏푢푠 + 휃1푏푢푠 ∙ 푧1푏푢푠 + 휃2 ∙ 푧2푏푢푠 + 휃3 ∙ 푧3푏푢푠 + 휃4 ∙ 푧4푏푢푠 + 휃5 ∙ 푧5푏푢푠 + 휃6 ∙ 푧6푏푢푠 + 휃7 ∙ 푧7푏푢푠

푉푟푎푖푙 = 휃2 ∙ 푧2푟푎푖푙 + 휃3 ∙ 푧3푟푎푖푙 + 휃4 ∙ 푧4푟푎푖푙 + 휃5 ∙ 푧5푟푎푖푙 + 휃6 ∙ 푧6푟푎푖푙 + 휃7 ∙ 푧7푟푎푖푙 (18) where

푥1푎푢푡표 = male⁡dummy⁡variable⁡times⁡automobile⁡dummy⁡variable

푥1푎푖푟 = male⁡dummy⁡variable⁡times⁡air⁡dummy⁡variable

푥1푏푢푠 = male⁡dummy⁡variable⁡times⁡bus⁡dummy⁡variable

푥2푎푢푡표 = dummy⁡variable⁡for⁡age⁡18 − 24⁡age⁡group⁡times⁡automobile⁡dummy⁡variable

푥2푎푖푟 = dummy⁡variable⁡for⁡age⁡18 − 24⁡age⁡group⁡times⁡air⁡dummy⁡variable

푥2푏푢푠 = dummy⁡variable⁡for⁡age⁡18 − 24⁡age⁡group⁡times⁡bus⁡dummy⁡variable

푥3푎푢푡표 = dummy⁡variable⁡for⁡age⁡70⁡or⁡older⁡age⁡group⁡times⁡automobile⁡dummy⁡variable

푥3푎푖푟 = dummy⁡variable⁡for⁡age⁡70⁡or⁡older⁡age⁡group⁡times⁡air⁡dummy⁡variable

푥3푏푢푠 = dummy⁡variable⁡for⁡age⁡70⁡or⁡older⁡age⁡group⁡times⁡bus⁡dummy⁡variable

푥4푎푢푡표 = income⁡times⁡automobile⁡dummy⁡variable

푥4푎푖푟 = income⁡times⁡air⁡dummy⁡variable

푥4푏푢푠 = income⁡times⁡bus⁡dummy⁡variable

푥5푎푢푡표 = dummy⁡variable⁡for⁡disability⁡times⁡automobile⁡dummy⁡variable⁡

푥5푎푖푟 = dummy⁡variable⁡for⁡disability⁡times⁡air⁡dummy⁡variable⁡

푥5푏푢푠 = dummy⁡variable⁡for⁡disability⁡times⁡bus⁡dummy⁡variable⁡

푦1푎푢푡표 = dummy⁡variable⁡for⁡traveling⁡alone⁡times⁡automobile⁡dummy⁡variable

푦1푎푖푟 = dummy⁡variable⁡for⁡traveling⁡alone⁡times⁡air⁡dummy⁡variable

푦1푏푢푠 = dummy⁡variable⁡for⁡traveling⁡alone⁡times⁡bus⁡dummy⁡variable

푦2푎푢푡표 = dummy⁡variable⁡for⁡personal⁡trip⁡times⁡automobile⁡dummy⁡variable

푦2푎푖푟 = dummy⁡variable⁡for⁡personal⁡trip⁡times⁡air⁡dummy⁡variable

푦2푏푢푠 = dummy⁡variable⁡for⁡personal⁡trip⁡times⁡bus⁡dummy⁡variable

푧1푎푢푡표 = automobile⁡dummy⁡variable

푧1푎푖푟 = air⁡dummy⁡variable

푧1푏푢푠 = bus⁡dummy⁡variable

푧2푎푢푡표 = automible⁡travel⁡time

푧2푎푖푟 = air⁡travel⁡time

푧2푏푢푠 = bus⁡travel⁡time

푧3푎푢푡표 = automobile⁡travel⁡cost

푧3푎푖푟 = air⁡travel⁡cost

푧3푏푢푠 = bus⁡travel⁡cost

푧4푎푖푟 = air⁡access⁡distance

푧4푏푢푠 = bus⁡access⁡distance

푧4푟푎푖푙 = rail⁡access⁡distance

푧5푎푖푟 = air⁡egress⁡distance

푧5푎푖푟 = bus⁡egress⁡distance

푧5푎푖푟 = rail⁡egress⁡distance

푧6푎푖푟 = air⁡frequency⁡dummy⁡variable⁡for⁡lowest⁡frequency

푧6푏푢푠 = bus⁡frequency⁡dummy⁡variable⁡for⁡lowest⁡frequency

푧6푟푎푖푙 = rail⁡frequency⁡dummy⁡variable⁡for⁡lowest⁡frequency

⁡푧7푎푖푟 = air⁡frequency⁡dummy⁡variable⁡for⁡highest⁡frequency

⁡푧7푏푢푠 = bus⁡frequency⁡dummy⁡variable⁡for⁡highest⁡frequency

⁡푧7푟푎푖푙 = rail⁡frequency⁡dummy⁡variable⁡for⁡highest⁡frequency

In this model, the characteristics of the individual included age, gender, income, and if they have a disability. Interaction terms between these variables and mode-specific dummy variables were included in the model to show the effect of individual characteristics on choice of specific modes.

Age was categorized into three groups: 18-24, 25-69, and 70 and older. Travel behavior may differ between these three age groups. The 18-24 age group, which consists of college students and younger adults, has been shown in some research to be more likely to use intercity bus services. The 25-69 age group consists of working-age individuals who, on average, drive more and may be more likely to choose automobile travel. The 70 and older group consists of older adults who may have reduced driving abilities and may be less inclined to drive long distances or travel by air. The model includes dummy variables for the 18-24 age group and 70+ age group, with the middle age group used as the reference.

Income was measured with a 1-5 scale, with 1 = less than $25,000, 2 = $25,000 to $49,999, 3 = $50,000 to $74,999, 4 = $75,000 to $99,999, and 5 = $100,000 or more. Disability was measured with a dummy variable equal to 1 if the individual answered that they have a medical condition or disability that makes it difficult to travel, and 0 otherwise.

Trip-specific characteristics included party size (traveling alone vs. in a group) and trip purpose (business or personal). Again, interaction terms between these variables and mode-specific dummy variables were included in the model to show the effect of trip characteristics on choice of specific modes.

Lastly, mode attributes included travel time, price, access distance, egress distance, and service frequency. Travel time was measured in hours, price in dollars per mile, access and egress distance in miles, and service frequency with dummy variables indicating the lowest and highest levels of frequency. Mode-specific dummy variables were also included in the model to account for mode preferences not accounted for by other variables in the model. The rail dummy variable was excluded from the model, so rail is considered the base or reference mode.

The model tests the following hypotheses, which were developed based on theory, findings from previous research, and survey results:

 Gender influences mode choice; men are more likely than women to drive.  Younger adults (aged 18-24) are more likely than others to choose intercity bus.  Older adults (aged 70 or older) are more likely than others to choose intercity bus.  Those with higher income are less likely to choose bus and more likely to choose automobile or air.  People with a disability are less likely to drive and more likely to choose bus.  People traveling alone are more likely to choose bus than those traveling in a group.  Those traveling for personal reasons are more likely to choose the automobile than those traveling for business.  Travel time negatively impacts choice of mode.  Travel cost negatively impacts choice of mode.  Access distance negatively impacts choice of mode.  Egress distance negatively impacts choice of mode.  Service frequency positively impacts choice of mode.  Everything else equal, travelers are less likely to choose bus than other modes.

Multinomial logit, nested logit, and mixed logit specifications of the model were all attempted. First, a multinomial logit model was estimated, and a Hausman specification test was conducted to analyze the IIA assumption (SAS 2010). Under the null hypothesis, the IIA property holds. Based on the test, which was run using the %IIA macro in SAS, the null hypothesis was rejected (p-value <0.01), indicating that an alternative model could be more appropriate.

Next, a nested logit model was estimated, as specified in section 3.3.3, with bus and rail nested together as an upper level choice. However, the logsum parameter for the public transit nest was not found to be statistically significant, indicating that the nest was not necessary.

A mixed logit model, therefore, was specified, based on it being a state-of-the-art model that does not rely on the IIA assumption. To select random parameters, following Hensher and Greene (2003), all parameters were first assumed to be random and then their estimated standard deviations were examined using a t-test. This approach was conducted, and two variables were found to have statistically significant standard deviations (travel time and price). These variables were then selected to be random in the final model. The random parameters were assumed to take a normal distribution.

The use of stated preference data with multiple SP responses from individual respondents results in repeated observations, or multiple observations from the same individual. Mehndiratta and Hansen (1997) noted that analysis of such data is complicated by the correlation of responses across the choices made by a single individual. They found that random parameter models offer advantages over standard logit models when repeated choice data are used. Further, Rose et al. (2011) argued that large improvements have been made with the use of the mixed multinomial logit model in cases where datasets contain multiple observations per respondent.

The analyses were conducted using SAS 9.3. Appendix C shows SAS code that was used, and SAS printouts of each of the models are provided in Appendix D. SAS code was developed with guidance from the SAS User’s Guide (SAS 2010). 5.2 Binary Logit Models to Estimate Effects of Attitudes on Mode Choice

In addition to the individual, trip, and mode characteristics included in the mode choice model, individuals have their own beliefs and attitudes about travel that are not directly observable but influence mode choice. As shown in section 4.2.4, the survey asked respondents a series of opinion questions regarding travel preferences and attitudes. These attitudinal variables were not included in the mixed logit model because doing so would have required adding numerous variables. Further, because attitudinal characteristics are not directly observable, including them in a mode choice model is not as useful for predicting ridership. Some of those attitudes may also be correlated to demographic characteristics already included in the model.

However, examining the influence of attitudes is important for understanding why some people choose a particular mode of travel. Knowledge about how attitudes influence mode choice can be used by intercity bus and rail companies for improving and marketing their services to attract new customers.

To examine the impact of attitudes on mode choice, separate binary logit models were estimated for each of the four modes. In these models, the dependent variable was equal to 1 if the mode was chosen and 0 if not. Mode characteristics (travel time and cost for the mode and competing modes) and trip characteristics (trip purpose, and party size) were included as explanatory variables with the addition of the 27 attitudinal responses given by the individual. These attitudinal variables were measures on a scale from 1 to 10, with a higher number indicating greater agreement with the attitude statement.

The binary logit models test the following hypotheses:

 Those with greater concern for the environment would be more likely to choose bus or rail travel and less likely to choose automobile travel.  Those who want to make more productive use or their time or who want to do something else while traveling would be less likely to choose auto and more likely to choose the alternatives.  Those who are more sensitive to travel time would be more likely to choose air and less likely to choose bus or rail.  Those who value flexibility would be more likely to travel by automobile.  Those concerned about privacy would be more likely to choose the automobile and less likely to choose air, bus, or rail.  Those concerned about safety and stress would be more likely to travel by bus or rail.  Individuals are more likely to choose a given mode if they perceive other users of that mode to be similar to them.

6. MODE CHOICE MODEL RESULTS

6.1 Mixed Logit Model Results

A total of 4,607 observations were used to estimate the model. Estimated results are shown in Table 6.1. The table also shows odds ratios for the statistically significant variables. The odds ratio is a way of comparing whether the probability of an event is the same for two groups of people. The odds of an event happening is equal to the probability of it happening divided by the probability of it not happening. An odds ratio is calculated by dividing the odds in group 1 by the odds in group 2. An odds ratio of 1 indicates the event is equally probable for the two groups, while an odds ratio greater than 1 indicates the event is more likely among the first group. An odds ratio less than 1 indicates the event is less likely among the first group. For variables not represented by a dummy variable, such as income (which is measured on a scale), travel time, price, and access and egress distance, the odds ratio is the estimated change in the odds of choosing a mode with a one unit increase in the variable.

Table 6.1 Estimated Mixed Logit Mode Choice Model Parameter Variable Variable description Estimate t value Odds Ratio Individual characteristics 푥1푎푢푡표 Male x Auto 0.4425 3.98*** 1.56 푥1푎푖푟 Male x Air 0.0871 0.46 푥1푏푢푠 Male x Bus 0.508 2.76*** 1.66 푥2푎푢푡표 Age 18-24 x Auto 0.1778 0.97 푥2푎푖푟 Age 18-24 x Air 0.123 0.39 푥2푏푢푠 Age 18-24 x Bus 0.4762 1.68* 1.61 푥3푎푢푡표 Age 70+ x Auto 0.135 0.73 푥3푎푖푟 Age 70+ x Air -0.7658 -2.29** 0.46 푥3푏푢푠 Age 70+ x Bus 0.0712 0.25 푥4푎푢푡표 Income x Auto 0.096 2.22** 1.10 푥4푎푖푟 Income x Air 0.1905 2.63*** 1.21 푥4푏푢푠 Income x Bus -0.1167 -1.57 푥5푎푢푡표 Disability x Auto -0.8052 -4.3*** 0.45 푥5푎푖푟 Disability x Air -1.4034 -3.16*** 0.25 푥5푏푢푠 Disability x Bus 0.2571 0.91 Trip characteristics 푦1푎푢푡표 Travel alone x Auto -0.3531 -3.28*** 0.70 푦1푎푖푟 Travel alone x Air -0.2013 -1.11 푦1푏푢푠 Travel alone x Bus 0.1236 0.69 푦2푎푢푡표 Personal trip x Auto 0.3108 2.88*** 1.36 푦2푎푖푟 Personal trip x Air -0.4977 -2.67*** 0.61 푦2푏푢푠 Personal trip x Bus -0.1997 -1.1 Mode characteristics 푧1푎푢푡표 Auto dummy 1.1581 5.87*** 3.18 푧1푎푖푟 Air dummy 0.9613 2.95*** 2.62 푧1푏푢푠 Bus dummy -0.8797 -2.9*** 0.41 푧2푚 Travel time mean -0.2706 -6.21*** 0.76 푧2푠 Travel time st. dev. -0.6322 -10.16*** 0.53 푧3푚 Price mean -5.4204 -11.04*** 0.00 푧3푠 Price st. dev. -2.4231 -8.49*** 0.09 푧4 Access distance -0.0189 -3.57*** 0.98 푧5 Egress distance -0.005461 -0.53 푧6 Frequency level 1 -0.0149 -0.16 푧7 Frequency level 3 0.0197 0.22 Goodness of fit measures Cragg-Uhler 1 0.7296 Adjusted Estrella 0.8253 Cragg-Uhler 2 0.7783 McFadden's LRI 0.4717 Veall- Estrella 0.8296 0.7711 Zimmermann *,**,*** denote significance at the 10%, 5%, and 1% levels, respectively

6.1.1 Impacts of Individual Characteristics

Among the individual characteristics, gender, age, income, and disability were all found to have some effect on mode choice. Men were found to have greater odds of choosing automobile or bus travel.

Regarding age, adults aged 70 or older were found to be less likely to choose air travel. Odds of choosing air travel decreased by 54% for older adults. Younger adults, aged 18-24, were found to be more likely to choose bus travel. The odds of choosing bus travel increased 61% for younger adults. This finding is consistent with other research showing younger adults being more likely to choose bus travel. The negative impact of age on choice of air travel was also found in previous research by Mattson et al. (2010a,b).

Income was found to have a statistically significant impact on mode choice, as expected. As income increased, individuals were more likely to choose automobile and air travel.

The impact of having a disability on mode choice was also found to be statistically significant and large in magnitude. People with a disability were much less likely to choose automobile or air travel and, therefore, more likely to choose bus and rail travel, compared to those without a disability.

6.1.2 Impacts of Trip Characteristics

The results show that party size and trip purpose have some significant effects on mode choice. Individuals traveling alone were found to be less likely to choose automobile travel, and, therefore, more likely to choose air, bus, or rail travel. In other words, respondents were more likely to choose automobile travel if they were traveling with a group of people. This result makes sense, because automobile travel becomes more cost effective as the size of the travel group increases (travel by other modes would require multiple fare payments).

Those taking personal trips were found to be more likely to choose automobile travel and less likely to choose air travel. The odds of choosing the automobile increased 36% if it was a personal trip, while the odds of choosing air decreased 39% for personal trips. Conversely, if the person was asked to take a business trip, the probability of choosing air increased and the probability of choosing automobile decreased. This result also makes sense, because business travelers may place a higher value on time and could be less sensitive to trip cost.

6.1.3 Impacts of Mode Characteristics

Travel time and travel price (consisting of fares and price of gasoline) were found to have negative and statistically significant impacts on mode choice, as expected. Access distances for bus, rail, and air travel were found to have negative and statistically significant impacts on mode choice, also as expected, but the impact of egress distance, while shown to also be negative, was not found to be statistically significant. Service frequency was not found to have a statistically significant effect.

The insignificant impacts of egress distance and service frequency may indicate that these variables are less important, or not important, but the complexity of the survey may have led respondents to focus on the factors they deemed most important, such as travel time and cost, while paying less attention to the other factors. In the real world, individuals may consider a wider range of variables. Another possible explanation for the insignificant impact of frequency is that individuals who do not typically use transit may not appreciate the importance of service frequency and, therefore, may not consider it when evaluating options.

Finally, the estimates for the mode-specific dummy variables show that, all other variables equal, respondents were significantly more likely to choose automobile travel (in comparison to all other modes) and air travel (in comparison to bus and rail), and were least likely to choose bus.

6.2 Attitudes and Mode Choice Results

Table 6.2 shows the odds ratios for each of the attitudes estimated from the binary logit model for each of the four modes. An odds ratio greater than 1 indicates an increased probability of choosing that mode if the respondent rated a higher agreement with the statement, and an odds ratio less than 1 indicates a decreased probability. (Estimated results for the mode and trip characteristics are not shown in the table but can be found in Appendix D.) A number of the results are statistically significant, as indicated by the asterisks. Many of the results are as expected, though some are unexpected.

The first three attitudes regard the environment. Respondents who gave higher scores tended to be more likely to choose bus and rail services and less likely to choose automobile and air, except for two results that were counter-intuitive: those who would be willing to pay more if it would help the environment were more likely to choose automobile and less likely to choose rail. The results suggest environmental concerns may play a role in supporting intercity bus and rail travel, but the results are inconsistent, and some results for bus travel are not statistically significant.

Those who would rather do something else with the time spent traveling were found to be less likely to choose automobile, which makes sense because unlike driving, other modes allow the individual to do other things while traveling. Similarly, those who would like to make productive use of their time when traveling were more likely to choose rail travel.

Those who placed a higher value on a predictable travel time were more likely to choose the automobile. This result suggests the other modes are not viewed as being as predictable.

Regarding travel time, those who would change their form of travel if it would save time were more likely to choose air travel, as expected. On the other hand, those more sensitive to time also had higher likelihoods of choosing bus and rail travel, though these effects were smaller in size.

Travelers who do not mind traveling with strangers were more likely to choose bus and rail travel and less likely to travel by automobile, which makes sense. Similarly, those who prefer to make trips alone were more likely to choose automobile and less likely to choose bus.

Those more concerned about getting into an accident were less likely to choose automobile and more likely to choose air travel, and those more concerned about having a stress-free trip than reaching their destination quickly were less likely to choose air travel and more likely to choose bus travel. These results suggest that the automobile is viewed as being less safe and air travel as safer. Further, bus travel is perceived as being less stressful. Intercity bus services could market their services to those more concerned about having a safe and stress-free trip.

Results showed that those who place a higher value on cleanliness were more likely to choose automobile travel, suggesting there may be negative perceptions regarding the cleanliness of the alternative modes, whether warranted or not.

Sensitivity to cost, as well as the value of convenience, was also shown to have an effect, with bus and rail users being more sensitive to cost. Those who were more likely to use the most convenient form of transportation, regardless of cost, were more likely to choose auto and air travel and less likely to choose bus or rail.

Lastly, the results suggest that individuals are more likely to choose a mode of travel if they feel that the people who use that mode are similar to them. Similar results were found by Mattson et al. (2010b). This effect was found to exist for intercity bus, rail, and air.

Table 6.2 Impacts of Attitudes on Mode Choice, Results from Binary Logit Model Opinion Statement Auto Air Bus Rail ------Odds Ratios------People who travel alone should pay more to help improve the 0.95** 0.97 1.01 1.11*** environment. I would be willing to pay more when I travel if it would help the 1.09*** 0.90*** 1.09* 0.93** environment. I would switch to a different form of transportation if it would help 0.92*** 1.05 1.02 1.10*** the environment. I would rather do something else with the time that I spend 0.94*** 1.13 0.95 1.02 traveling. I would like to make productive use of my time when traveling. 0.99 0.93* 1.06 1.09** I prefer a travel option that has a predictable travel time. 1.08** 0.97 0.91 0.93 When traveling, I like to keep as close as possible to my departure 1.01 1.00 0.97 1.03 and arrival schedules. If my travel options are delayed, I want to know the cause and 1.00 0.88*** 1.06 1.09* length of the delay. I would change my form of travel if it would save me some time. 0.87*** 1.19*** 1.10** 1.06* I always take the fastest route to my destination even if I have a 0.98 0.97 1.03 1.05* cheaper alternative. I don’t mind traveling with strangers. 0.92*** 1.00 1.13*** 1.08*** When traveling, I like to talk and visit with other people. 1.10*** 0.98 0.96 0.87*** I prefer to make trips alone, because I like the time to myself. 1.04* 0.99 0.91** 0.97 I worry about getting in an accident when I travel. 0.95*** 1.06** 1.05 1.00 Having privacy is important to me when I travel. 1.03 1.00 0.97 0.97 I need to make trips according to a fixed schedule. 0.98 1.06** 1.03 0.96 It’s important to be able to change my travel plans at a moment’s 1.04** 0.98 0.97 0.99 notice. Having a stress-free trip is more important than reaching my 1.02 0.87*** 1.14** 1.05 destination quickly. I don’t mind long delays as long as I’m comfortable. 0.94*** 1.08*** 0.91** 1.08*** It is important to have comfortable seats when I travel. 0.99 1.02 0.96 1.01 I avoid traveling at certain times because it is too stressful. 1.01 0.98 1.03 1.00 A clean vehicle is important to me. 1.08*** 0.92*** 0.96 0.94** I use the most convenient form of transportation regardless of cost. 1.04** 1.05** 0.93** 0.91*** The people who ride intercity bus are like me. 0.99 1.02 1.12* 0.98 The people who fly are like me. 1.03 1.09*** 0.90** 0.88*** The people who use intercity rail service are like me. 0.89*** 0.95 0.98 1.33*** I’m willing to pay more for a ticket if it allows me to re-book my 0.97* 1.03 1.09** 0.99 trip later for free. *,**,*** denote significance at the 10%, 5%, and 1% levels, respectively

7. APPLYING THE RESULTS FROM THE MODE CHOICE MODEL

Results from the mixed logit model can be used to estimate mode shares between origins and destinations, and it can be incorporated into a statewide travel demand model to estimate mode volumes. Section 7.1 provides a description of a statewide travel demand model, and Section 7.2 describes how mode shares were estimated for each origin-destination (OD) pair. To help understand how the results from the mode choice model can be applied to estimate mode shares, Section 7.3 provides an example calculation. A number of alternative scenarios were then analyzed, with the results presented in Section 7.4.

7.1 North Dakota Statewide Passenger Travel Demand Model

North Dakota does not currently have a statewide travel demand model for passenger travel, but this study shows how the mode choice results could be incorporated into such a model. The state was divided into traffic analysis zones (TAZs) based on census block groups. The model has 572 TAZs in North Dakota, following the 572 census block groups in the state, and 96 out-of-state TAZs, for a total of 668 TAZs, as shown in Figure 7.1.

Figure 7.1 Traffic Analysis Zones for North Dakota Travel Demand Model

A road network for the state and external TAZs was developed consisting of all U.S., state, and local roadways. Speed limits were also estimated for each roadway segment to estimate travel times between points. Next, the intercity bus and rail networks were created, as shown in Figures 7.2 and 7.3.

Figure 7.2 North Dakota Intercity Bus Network, February 2016

Figure 7.3 North Dakota Intercity Rail Network

The intercity bus network depicted in Figure 7.2 was the existing network as of February 2016. There have been regular changes to the network, including changes later in 2016, so this network is not current as of publication. However, for the sake of this study, it will be referred to as the current network. The objective of this demonstration is to show how the results from this study can be used with a given intercity bus network. The network shown in Figure 7.2 includes nine bus stops in North Dakota, including two different stops in Grand Forks, and stops in Sidney, Montana, and Sisseton, South Dakota. The Sidney and Sisseton stops were included because they are the closest bus stops for some North Dakota residents.

The bus network consists of three routes (with service in both directions), all operated by Jefferson Lines. Each route terminates or originates in Fargo. One route travels east-west across the state connecting Fargo to Williston, by way of Bismarck and Minot. The route also travels into Montana, on its way to Billings. Two north-south routes are also provided (in addition to the north-south service between Bismarck and Minot on the previously described route). One provides service between Fargo and Sioux Falls, South Dakota, and the other provides service between Fargo and Grand Forks.

Although each route connects to Fargo, each route provides just one trip per day, and the schedules, in many cases, do not provide meaningful connections that would allow for same-day transfers. For example, service from Grand Forks to Fargo arrives in Fargo at 9:00 p.m., but services that depart Fargo to the west or south leave much earlier in the day, so residents in Grand Forks cannot travel elsewhere on the network without spending a night in Fargo. Similarly, service into Fargo from South Dakota arrives late at night. The current schedules allow for one relatively easy transfer. Service from the west into Fargo arrives at 11:55 a.m., while service into South Dakota departs Fargo at 1:00 p.m. These routes, therefore, were considered to be connected in the network, while the other routes were considered to be disconnected. When estimating travel times by bus between different origins and destinations, it was assumed that transfers would not be made between the disconnected routes.

The intercity rail network consists of one route connecting Fargo and Williston, with service provided once per day in each direction. This is part of Amtrak’s Empire Builder route, which provides service between Seattle and Chicago.

Bus and rail travel times between stops were calculated based on schedules given on the Jefferson Lines and Amtrak websites as of February 2016, as shown in Tables 7.1 and 7.2.

Table 7.1 Travel Time between Intercity Bus Stops Destination

Grand Valley Origin Forks Fargo City Jamestown Bismarck Minot Stanley Williston Sidney Sisseton ------minutes------Grand Forks 0 85 ------Fargo 80 0 65 105 210 315 415 490 566 65 Valley City - 55 0 40 145 250 350 425 501 185 Jamestown - 105 35 0 105 210 310 385 461 235 Bismarck - 210 140 105 0 105 205 280 356 340 Minot - 330 260 225 105 0 75 150 226 460 Stanley - 450 380 345 225 65 0 75 151 580 Williston - 530 450 425 305 145 80 0 61 660 Sidney - 605 525 500 380 220 155 60 0 735 Sisseton - 75 ------

Table 7.2 Travel Time between Intercity Rail Stations Destination Grand Devils Origin Fargo Forks Lake Rugby Minot Stanley Williston ------minutes------Fargo 0 77 158 212 299 387 457 Grand Forks 76 0 81 135 222 310 380 Devils Lake 161 85 0 54 141 229 299 Rugby 215 139 54 0 87 175 245 Minot 271 195 110 56 0 51 121 Stanley 362 286 201 147 71 0 70 Williston 424 348 263 209 133 62 0

Travel times between each pair of TAZs were calculated for automobile, bus, and rail travel based on road, bus, and rail networks. Centroids were created for each TAZ, and travel times between centroids were estimated.

First, travel times between TAZ centroids for automobile travel were estimated by finding paths with the shortest travel time (based on travel distance and speed). Bus and rail travel included both in-vehicle travel time and access and egress travel. For each O-D pair, access distance was found by calculating the distance on the road network to the nearest bus or rail station. Similarly, egress distance was calculated as the distance on the road network from the final destination to the bus or rail station closest to the final destination. In-vehicle travel times were then calculated based on the travel times shown in Tables 6.1 and 6.2. These calculations were made for every O-D pair, though in many cases, bus or rail travel would be impractical or unrealistic because of very long access or egress distances.

7.2 Estimating Mode Shares

Mode shares for automobile, bus, and rail travel were calculated for each O-D pair using the calculated travel times for each mode, access distances for bus and rail, given gas prices and fares for bus and rail, income levels for the origin zone, and the percentage of population in the origin zone aged 18-24. Air travel was not considered in this analysis because in-state intercity trips by air are rare given current service levels.

Average income levels and percentage of population aged 18-24 for each origin TAZ were calculated using data from the ACS. Because the ACS does not include disability data at the block group level, the analysis assumed that in each TAZ, 6% of the population has a medical condition or disability that makes it difficult to travel. This estimate is based on the results from the survey. It was also assumed that travelers were evenly split between males and females in each zone.

Separate mode shares were calculated for personal trips and business trips. It was assumed that most business trips would be made alone, and that personal trips would include a greater percentage of group trips. Data from the 2009 National Household Travel Survey showed that the average vehicle occupancy was 1.13 for work trips, 1.78 for shopping trips, and 2.20 for social and recreational trips (Santos et al. 2011). A rough estimate was made that 90% of business trips and 20% of personal trips would be made alone.

While the mode choice model did not show egress distance having a significant effect on mode share, it can reasonably be expected that bus and rail are not an option when access or egress distance is greater than a certain distance. In this analysis, is assumed that trips would not be made by bus or rail if access or egress distance is greater than 25 miles. Previous research tends to suggest catchment areas of 10 to 20 miles, though it could be somewhat higher in North Dakota, where travel distances are generally greater. Figures 7.4 and 7.5 highlight TAZs within 25 miles of a bus stop or rail station.

Figure 7.4 TAZs within 25 Miles of Intercity Bus Stop

Figure 7.5 TAZs within 25 Miles of Intercity Rail Station

Mode shares were then estimated for each OD pair. Because bus and rail are not an option for many trips, due to long access or egress distance, the automobile mode share is 100% for many OD pairs. Some OD pairs also represents trips that are not intercity, since most cities contain multiple TAZs, and intercity bus and rail are therefore not an option.

To illustrate intercity mode shares between the main cities across the state, the 30 largest cities or metro areas were identified, and a matrix of mode shares was calculated for these cities. Figure 7.6 shows these cities and the TAZs that intersect or are contained within them. A significant majority of intercity trips originate or terminate in these TAZs. Because most of these cities contain multiple TAZs, calculating mode shares between cities required taking a weighted average of mode shares between TAZs based on the number of households in each TAZ. The cites of Fargo, West Fargo, and Horace were combined into one metro area, and the cities of Bismarck, Mandan, and Lincoln were similarly combined.

Figure 7.6 Most Populated Cities in North Dakota and Corresponding TAZs

Table 7.3 shows the estimated bus mode shares between these 30 cities/metro areas for personal trips, and Table 7.4 shows the same for business trips. These estimates are based on a gas price of $2.00 per gallon and intercity bus and rail fares at $0.20 per mile. Among these 30 cities, only nine are served by any intercity bus trips. Eight of these cities are served by a bus station, and the ninth, Casselton, is within 25 miles of the Fargo bus station. The five largest cities or metro areas are all served by intercity bus. Among the 10 largest, Dickinson, Wahpeton, and Devils Lake are not served by intercity bus as of the time of this analysis (recent changes have added service to Dickinson).

Table 7.3 Estimated Bus Mode Shares for Personal Trips

Destination

→

Minot 2.9 4.4 0.0 - 3.9 0.0 2.6 0.0 0.0 2.8 0.0 0.0 0.0 0.0 2.7 0.0 4.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Williston 1.8 2.0 0.0 3.7 - 0.0 1.7 0.0 0.0 1.9 0.0 0.0 0.0 0.0 1.7 0.0 4.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Dickinson 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Jamestown 4.0 4.2 0.0 3.5 2.6 0.0 - 0.0 0.0 4.5 0.0 0.0 0.0 0.0 3.8 0.0 2.7 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Wahpeton 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Devils Lake 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Valley City 4.2 3.7 0.0 3.1 2.3 0.0 4.1 0.0 0.0 - 0.0 0.0 0.0 0.0 3.9 0.0 2.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Grafton 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Watford City 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Beulah 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Rugby 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Casselton 0.0 2.2 2.9 2.0 1.4 0.0 2.4 0.0 0.0 2.8 0.0 0.0 0.0 0.0 - 0.0 1.5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Hazen 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Stanley 2.0 2.3 0.0 4.0 4.1 0.0 1.8 0.0 0.0 2.0 0.0 0.0 0.0 0.0 1.8 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

New Town 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Bottineau 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Lisbon 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Carrington 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Mayville 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Langdon 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Oaks 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0

Harvey 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0

Bowman 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0

Tioga 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0

Hillsboro 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0

Garrison 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - 0.0 New 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 - Rockford

Table 7.4 Estimated Bus Mode Shares for Business Trips

Destination

→

Origin

Fargo Bismarck Grand Forks Minot Williston Dickinson Jamestown Wahpeton DevilsLake CityValley Grafton Watford City Beulah Rugby Casselton Hazen Stanley TownNew Bottineau Lisbon Carrington Mayville Langdon Oaks Harvey Bowman Tioga Hillsboro Garrison New Rockford

Fargo - 5.1 6.9 5.2 3.8 0.0 5.4 0.0 0.0 7.3 0.0 0.0 0.0 0.0 0.0 0.0 4.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 Bismarck 5.8 - 0.0 7.0 5.0 0.0 6.7 0.0 0.0 6.5 0.0 0.0 0.0 0.0 5.5 0.0 5.6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 Grand Forks 7.7 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 7.9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Minot 4.9 7.5 0.0 - 6.5 0.0 3.9 0.0 0.0 4.4 0.0 0.0 0.0 0.0 4.7 0.0 6.9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Williston 3.1 3.0 0.0 6.2 - 0.0 2.6 0.0 0.0 3.0 0.0 0.0 0.0 0.0 2.9 0.0 6.7 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Dickinson 0.0 0.0 0.0 0.0 0.0 - 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Jamestown 6.8 7.1 0.0 5.9 4.4 0.0 - 0.0 0.0 7.6 0.0 0.0 0.0 0.0 6.4 0.0 4.7 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0